4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
21 #include <linux/module.h>
22 #include <linux/nmi.h>
23 #include <linux/init.h>
24 #include <asm/uaccess.h>
25 #include <linux/highmem.h>
26 #include <linux/smp_lock.h>
27 #include <linux/pagemap.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/completion.h>
31 #include <linux/kernel_stat.h>
32 #include <linux/security.h>
33 #include <linux/notifier.h>
34 #include <linux/suspend.h>
35 #include <linux/blkdev.h>
36 #include <linux/delay.h>
37 #include <linux/smp.h>
38 #include <linux/timer.h>
39 #include <linux/rcupdate.h>
40 #include <linux/cpu.h>
41 #include <linux/percpu.h>
42 #include <linux/kthread.h>
43 #include <linux/vserver/sched.h>
44 #include <linux/vs_base.h>
47 #include <asm/unistd.h>
50 #define cpu_to_node_mask(cpu) node_to_cpumask(cpu_to_node(cpu))
52 #define cpu_to_node_mask(cpu) (cpu_online_map)
56 * Convert user-nice values [ -20 ... 0 ... 19 ]
57 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
60 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
61 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
62 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
65 * 'User priority' is the nice value converted to something we
66 * can work with better when scaling various scheduler parameters,
67 * it's a [ 0 ... 39 ] range.
69 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
70 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
71 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
72 #define AVG_TIMESLICE (MIN_TIMESLICE + ((MAX_TIMESLICE - MIN_TIMESLICE) *\
73 (MAX_PRIO-1-NICE_TO_PRIO(0))/(MAX_USER_PRIO - 1)))
76 * Some helpers for converting nanosecond timing to jiffy resolution
78 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
79 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
82 * These are the 'tuning knobs' of the scheduler:
84 * Minimum timeslice is 10 msecs, default timeslice is 100 msecs,
85 * maximum timeslice is 200 msecs. Timeslices get refilled after
88 #define MIN_TIMESLICE ( 10 * HZ / 1000)
89 #define MAX_TIMESLICE (200 * HZ / 1000)
90 #define ON_RUNQUEUE_WEIGHT 30
91 #define CHILD_PENALTY 95
92 #define PARENT_PENALTY 100
94 #define PRIO_BONUS_RATIO 25
95 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
96 #define INTERACTIVE_DELTA 2
97 #define MAX_SLEEP_AVG (AVG_TIMESLICE * MAX_BONUS)
98 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
99 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
100 #define CREDIT_LIMIT 100
103 * If a task is 'interactive' then we reinsert it in the active
104 * array after it has expired its current timeslice. (it will not
105 * continue to run immediately, it will still roundrobin with
106 * other interactive tasks.)
108 * This part scales the interactivity limit depending on niceness.
110 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
111 * Here are a few examples of different nice levels:
113 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
114 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
115 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
116 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
117 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
119 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
120 * priority range a task can explore, a value of '1' means the
121 * task is rated interactive.)
123 * Ie. nice +19 tasks can never get 'interactive' enough to be
124 * reinserted into the active array. And only heavily CPU-hog nice -20
125 * tasks will be expired. Default nice 0 tasks are somewhere between,
126 * it takes some effort for them to get interactive, but it's not
130 #define CURRENT_BONUS(p) \
131 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
135 #define TIMESLICE_GRANULARITY(p) (MIN_TIMESLICE * \
136 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
139 #define TIMESLICE_GRANULARITY(p) (MIN_TIMESLICE * \
140 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
143 #define SCALE(v1,v1_max,v2_max) \
144 (v1) * (v2_max) / (v1_max)
147 (SCALE(TASK_NICE(p), 40, MAX_BONUS) + INTERACTIVE_DELTA)
149 #define TASK_INTERACTIVE(p) \
150 ((p)->prio <= (p)->static_prio - DELTA(p))
152 #define INTERACTIVE_SLEEP(p) \
153 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
154 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
156 #define HIGH_CREDIT(p) \
157 ((p)->interactive_credit > CREDIT_LIMIT)
159 #define LOW_CREDIT(p) \
160 ((p)->interactive_credit < -CREDIT_LIMIT)
163 * BASE_TIMESLICE scales user-nice values [ -20 ... 19 ]
164 * to time slice values.
166 * The higher a thread's priority, the bigger timeslices
167 * it gets during one round of execution. But even the lowest
168 * priority thread gets MIN_TIMESLICE worth of execution time.
170 * task_timeslice() is the interface that is used by the scheduler.
173 #define BASE_TIMESLICE(p) (MIN_TIMESLICE + \
174 ((MAX_TIMESLICE - MIN_TIMESLICE) * \
175 (MAX_PRIO-1 - (p)->static_prio) / (MAX_USER_PRIO-1)))
177 static unsigned int task_timeslice(task_t *p)
179 return BASE_TIMESLICE(p);
182 #define task_hot(p, now, sd) ((now) - (p)->timestamp < (sd)->cache_hot_time)
185 * These are the runqueue data structures:
187 typedef struct runqueue runqueue_t;
189 #ifdef CONFIG_CKRM_CPU_SCHEDULE
190 #include <linux/ckrm_classqueue.h>
193 #ifdef CONFIG_CKRM_CPU_SCHEDULE
196 * if belong to different class, compare class priority
197 * otherwise compare task priority
199 #define TASK_PREEMPTS_CURR(p, rq) \
200 (((p)->cpu_class != (rq)->curr->cpu_class) && ((rq)->curr != (rq)->idle))? class_preempts_curr((p),(rq)->curr) : ((p)->prio < (rq)->curr->prio)
202 #define BITMAP_SIZE ((((MAX_PRIO+1+7)/8)+sizeof(long)-1)/sizeof(long))
204 unsigned int nr_active;
205 unsigned long bitmap[BITMAP_SIZE];
206 struct list_head queue[MAX_PRIO];
208 #define rq_active(p,rq) (rq->active)
209 #define rq_expired(p,rq) (rq->expired)
210 #define ckrm_rebalance_tick(j,this_cpu) do {} while (0)
211 #define TASK_PREEMPTS_CURR(p, rq) \
212 ((p)->prio < (rq)->curr->prio)
216 * This is the main, per-CPU runqueue data structure.
218 * Locking rule: those places that want to lock multiple runqueues
219 * (such as the load balancing or the thread migration code), lock
220 * acquire operations must be ordered by ascending &runqueue.
226 * nr_running and cpu_load should be in the same cacheline because
227 * remote CPUs use both these fields when doing load calculation.
229 unsigned long nr_running;
230 #if defined(CONFIG_SMP)
231 unsigned long cpu_load;
233 unsigned long long nr_switches, nr_preempt;
234 unsigned long expired_timestamp, nr_uninterruptible;
235 unsigned long long timestamp_last_tick;
237 struct mm_struct *prev_mm;
238 #ifdef CONFIG_CKRM_CPU_SCHEDULE
239 unsigned long ckrm_cpu_load;
240 struct classqueue_struct classqueue;
242 prio_array_t *active, *expired, arrays[2];
244 int best_expired_prio;
248 struct sched_domain *sd;
250 /* For active balancing */
254 task_t *migration_thread;
255 struct list_head migration_queue;
257 struct list_head hold_queue;
261 static DEFINE_PER_CPU(struct runqueue, runqueues);
263 #define for_each_domain(cpu, domain) \
264 for (domain = cpu_rq(cpu)->sd; domain; domain = domain->parent)
266 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
267 #define this_rq() (&__get_cpu_var(runqueues))
268 #define task_rq(p) cpu_rq(task_cpu(p))
269 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
272 * Default context-switch locking:
274 #ifndef prepare_arch_switch
275 # define prepare_arch_switch(rq, next) do { } while (0)
276 # define finish_arch_switch(rq, next) spin_unlock_irq(&(rq)->lock)
277 # define task_running(rq, p) ((rq)->curr == (p))
280 #ifdef CONFIG_CKRM_CPU_SCHEDULE
281 #include <linux/ckrm_sched.h>
282 spinlock_t cvt_lock = SPIN_LOCK_UNLOCKED;
283 rwlock_t class_list_lock = RW_LOCK_UNLOCKED;
284 LIST_HEAD(active_cpu_classes); // list of active cpu classes; anchor
285 struct ckrm_cpu_class default_cpu_class_obj;
288 * the minimum CVT allowed is the base_cvt
289 * otherwise, it will starve others
291 CVT_t get_min_cvt(int cpu)
294 struct ckrm_local_runqueue * lrq;
297 node = classqueue_get_head(bpt_queue(cpu));
298 lrq = (node) ? class_list_entry(node) : NULL;
301 min_cvt = lrq->local_cvt;
309 * update the classueue base for all the runqueues
310 * TODO: we can only update half of the min_base to solve the movebackward issue
312 static inline void check_update_class_base(int this_cpu) {
313 unsigned long min_base = 0xFFFFFFFF;
317 if (! cpu_online(this_cpu)) return;
320 * find the min_base across all the processors
322 for_each_online_cpu(i) {
324 * I should change it to directly use bpt->base
326 node = classqueue_get_head(bpt_queue(i));
327 if (node && node->prio < min_base) {
328 min_base = node->prio;
331 if (min_base != 0xFFFFFFFF)
332 classqueue_update_base(bpt_queue(this_cpu),min_base);
335 static inline void ckrm_rebalance_tick(int j,int this_cpu)
337 #ifdef CONFIG_CKRM_CPU_SCHEDULE
338 read_lock(&class_list_lock);
339 if (!(j % CVT_UPDATE_TICK))
340 update_global_cvts(this_cpu);
342 #define CKRM_BASE_UPDATE_RATE 400
343 if (! (jiffies % CKRM_BASE_UPDATE_RATE))
344 check_update_class_base(this_cpu);
346 read_unlock(&class_list_lock);
350 static inline struct ckrm_local_runqueue *rq_get_next_class(struct runqueue *rq)
352 cq_node_t *node = classqueue_get_head(&rq->classqueue);
353 return ((node) ? class_list_entry(node) : NULL);
356 static inline struct task_struct * rq_get_next_task(struct runqueue* rq)
359 struct task_struct *next;
360 struct ckrm_local_runqueue *queue;
361 int cpu = smp_processor_id();
365 if ((queue = rq_get_next_class(rq))) {
366 array = queue->active;
367 //check switch active/expired queue
368 if (unlikely(!queue->active->nr_active)) {
369 queue->active = queue->expired;
370 queue->expired = array;
371 queue->expired_timestamp = 0;
373 if (queue->active->nr_active)
374 set_top_priority(queue,
375 find_first_bit(queue->active->bitmap, MAX_PRIO));
377 classqueue_dequeue(queue->classqueue,
378 &queue->classqueue_linkobj);
379 cpu_demand_event(get_rq_local_stat(queue,cpu),CPU_DEMAND_DEQUEUE,0);
382 goto retry_next_class;
384 BUG_ON(!queue->active->nr_active);
385 next = task_list_entry(array->queue[queue->top_priority].next);
390 static inline void rq_load_inc(runqueue_t *rq, struct task_struct *p) { rq->ckrm_cpu_load += cpu_class_weight(p->cpu_class); }
391 static inline void rq_load_dec(runqueue_t *rq, struct task_struct *p) { rq->ckrm_cpu_load -= cpu_class_weight(p->cpu_class); }
393 #else /*CONFIG_CKRM_CPU_SCHEDULE*/
395 static inline struct task_struct * rq_get_next_task(struct runqueue* rq)
398 struct list_head *queue;
402 if (unlikely(!array->nr_active)) {
404 * Switch the active and expired arrays.
406 rq->active = rq->expired;
409 rq->expired_timestamp = 0;
410 rq->best_expired_prio = MAX_PRIO;
413 idx = sched_find_first_bit(array->bitmap);
414 queue = array->queue + idx;
415 return list_entry(queue->next, task_t, run_list);
418 static inline void class_enqueue_task(struct task_struct* p, prio_array_t *array) { }
419 static inline void class_dequeue_task(struct task_struct* p, prio_array_t *array) { }
420 static inline void init_cpu_classes(void) { }
421 static inline void rq_load_inc(runqueue_t *rq, struct task_struct *p) { }
422 static inline void rq_load_dec(runqueue_t *rq, struct task_struct *p) { }
423 #endif /* CONFIG_CKRM_CPU_SCHEDULE */
427 * task_rq_lock - lock the runqueue a given task resides on and disable
428 * interrupts. Note the ordering: we can safely lookup the task_rq without
429 * explicitly disabling preemption.
431 runqueue_t *task_rq_lock(task_t *p, unsigned long *flags)
436 local_irq_save(*flags);
438 spin_lock(&rq->lock);
439 if (unlikely(rq != task_rq(p))) {
440 spin_unlock_irqrestore(&rq->lock, *flags);
441 goto repeat_lock_task;
446 void task_rq_unlock(runqueue_t *rq, unsigned long *flags)
448 spin_unlock_irqrestore(&rq->lock, *flags);
452 * rq_lock - lock a given runqueue and disable interrupts.
454 static runqueue_t *this_rq_lock(void)
460 spin_lock(&rq->lock);
465 static inline void rq_unlock(runqueue_t *rq)
467 spin_unlock_irq(&rq->lock);
471 * Adding/removing a task to/from a priority array:
473 void dequeue_task(struct task_struct *p, prio_array_t *array)
477 list_del(&p->run_list);
478 if (list_empty(array->queue + p->prio))
479 __clear_bit(p->prio, array->bitmap);
480 class_dequeue_task(p,array);
483 void enqueue_task(struct task_struct *p, prio_array_t *array)
485 list_add_tail(&p->run_list, array->queue + p->prio);
486 __set_bit(p->prio, array->bitmap);
489 class_enqueue_task(p,array);
493 * Used by the migration code - we pull tasks from the head of the
494 * remote queue so we want these tasks to show up at the head of the
497 static inline void enqueue_task_head(struct task_struct *p, prio_array_t *array)
499 list_add(&p->run_list, array->queue + p->prio);
500 __set_bit(p->prio, array->bitmap);
503 class_enqueue_task(p,array);
507 * effective_prio - return the priority that is based on the static
508 * priority but is modified by bonuses/penalties.
510 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
511 * into the -5 ... 0 ... +5 bonus/penalty range.
513 * We use 25% of the full 0...39 priority range so that:
515 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
516 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
518 * Both properties are important to certain workloads.
520 static int effective_prio(task_t *p)
527 bonus = CURRENT_BONUS(p) - MAX_BONUS / 2;
529 prio = p->static_prio - bonus;
530 if (__vx_task_flags(p, VXF_SCHED_PRIO, 0))
531 prio += effective_vavavoom(p, MAX_USER_PRIO);
533 if (prio < MAX_RT_PRIO)
535 if (prio > MAX_PRIO-1)
541 * __activate_task - move a task to the runqueue.
543 static inline void __activate_task(task_t *p, runqueue_t *rq)
545 enqueue_task(p, rq_active(p,rq));
551 * __activate_idle_task - move idle task to the _front_ of runqueue.
553 static inline void __activate_idle_task(task_t *p, runqueue_t *rq)
555 enqueue_task_head(p, rq_active(p,rq));
560 static void recalc_task_prio(task_t *p, unsigned long long now)
562 unsigned long long __sleep_time = now - p->timestamp;
563 unsigned long sleep_time;
565 if (__sleep_time > NS_MAX_SLEEP_AVG)
566 sleep_time = NS_MAX_SLEEP_AVG;
568 sleep_time = (unsigned long)__sleep_time;
570 if (likely(sleep_time > 0)) {
572 * User tasks that sleep a long time are categorised as
573 * idle and will get just interactive status to stay active &
574 * prevent them suddenly becoming cpu hogs and starving
577 if (p->mm && p->activated != -1 &&
578 sleep_time > INTERACTIVE_SLEEP(p)) {
579 p->sleep_avg = JIFFIES_TO_NS(MAX_SLEEP_AVG -
582 p->interactive_credit++;
585 * The lower the sleep avg a task has the more
586 * rapidly it will rise with sleep time.
588 sleep_time *= (MAX_BONUS - CURRENT_BONUS(p)) ? : 1;
591 * Tasks with low interactive_credit are limited to
592 * one timeslice worth of sleep avg bonus.
595 sleep_time > JIFFIES_TO_NS(task_timeslice(p)))
596 sleep_time = JIFFIES_TO_NS(task_timeslice(p));
599 * Non high_credit tasks waking from uninterruptible
600 * sleep are limited in their sleep_avg rise as they
601 * are likely to be cpu hogs waiting on I/O
603 if (p->activated == -1 && !HIGH_CREDIT(p) && p->mm) {
604 if (p->sleep_avg >= INTERACTIVE_SLEEP(p))
606 else if (p->sleep_avg + sleep_time >=
607 INTERACTIVE_SLEEP(p)) {
608 p->sleep_avg = INTERACTIVE_SLEEP(p);
614 * This code gives a bonus to interactive tasks.
616 * The boost works by updating the 'average sleep time'
617 * value here, based on ->timestamp. The more time a
618 * task spends sleeping, the higher the average gets -
619 * and the higher the priority boost gets as well.
621 p->sleep_avg += sleep_time;
623 if (p->sleep_avg > NS_MAX_SLEEP_AVG) {
624 p->sleep_avg = NS_MAX_SLEEP_AVG;
626 p->interactive_credit++;
631 p->prio = effective_prio(p);
635 * activate_task - move a task to the runqueue and do priority recalculation
637 * Update all the scheduling statistics stuff. (sleep average
638 * calculation, priority modifiers, etc.)
640 static void activate_task(task_t *p, runqueue_t *rq, int local)
642 unsigned long long now;
647 /* Compensate for drifting sched_clock */
648 runqueue_t *this_rq = this_rq();
649 now = (now - this_rq->timestamp_last_tick)
650 + rq->timestamp_last_tick;
654 recalc_task_prio(p, now);
657 * This checks to make sure it's not an uninterruptible task
658 * that is now waking up.
662 * Tasks which were woken up by interrupts (ie. hw events)
663 * are most likely of interactive nature. So we give them
664 * the credit of extending their sleep time to the period
665 * of time they spend on the runqueue, waiting for execution
666 * on a CPU, first time around:
672 * Normal first-time wakeups get a credit too for
673 * on-runqueue time, but it will be weighted down:
680 __activate_task(p, rq);
684 * deactivate_task - remove a task from the runqueue.
686 static void deactivate_task(struct task_struct *p, runqueue_t *rq)
690 if (p->state == TASK_UNINTERRUPTIBLE)
691 rq->nr_uninterruptible++;
692 dequeue_task(p, p->array);
697 * resched_task - mark a task 'to be rescheduled now'.
699 * On UP this means the setting of the need_resched flag, on SMP it
700 * might also involve a cross-CPU call to trigger the scheduler on
704 static void resched_task(task_t *p)
706 int need_resched, nrpolling;
709 /* minimise the chance of sending an interrupt to poll_idle() */
710 nrpolling = test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
711 need_resched = test_and_set_tsk_thread_flag(p,TIF_NEED_RESCHED);
712 nrpolling |= test_tsk_thread_flag(p,TIF_POLLING_NRFLAG);
714 if (!need_resched && !nrpolling && (task_cpu(p) != smp_processor_id()))
715 smp_send_reschedule(task_cpu(p));
719 static inline void resched_task(task_t *p)
721 set_tsk_need_resched(p);
726 * task_curr - is this task currently executing on a CPU?
727 * @p: the task in question.
729 inline int task_curr(const task_t *p)
731 return cpu_curr(task_cpu(p)) == p;
741 struct list_head list;
742 enum request_type type;
744 /* For REQ_MOVE_TASK */
748 /* For REQ_SET_DOMAIN */
749 struct sched_domain *sd;
751 struct completion done;
755 * The task's runqueue lock must be held.
756 * Returns true if you have to wait for migration thread.
758 static int migrate_task(task_t *p, int dest_cpu, migration_req_t *req)
760 runqueue_t *rq = task_rq(p);
763 * If the task is not on a runqueue (and not running), then
764 * it is sufficient to simply update the task's cpu field.
766 if (!p->array && !task_running(rq, p)) {
767 set_task_cpu(p, dest_cpu);
771 init_completion(&req->done);
772 req->type = REQ_MOVE_TASK;
774 req->dest_cpu = dest_cpu;
775 list_add(&req->list, &rq->migration_queue);
780 * wait_task_inactive - wait for a thread to unschedule.
782 * The caller must ensure that the task *will* unschedule sometime soon,
783 * else this function might spin for a *long* time. This function can't
784 * be called with interrupts off, or it may introduce deadlock with
785 * smp_call_function() if an IPI is sent by the same process we are
786 * waiting to become inactive.
788 void wait_task_inactive(task_t * p)
795 rq = task_rq_lock(p, &flags);
796 /* Must be off runqueue entirely, not preempted. */
797 if (unlikely(p->array)) {
798 /* If it's preempted, we yield. It could be a while. */
799 preempted = !task_running(rq, p);
800 task_rq_unlock(rq, &flags);
806 task_rq_unlock(rq, &flags);
810 * kick_process - kick a running thread to enter/exit the kernel
811 * @p: the to-be-kicked thread
813 * Cause a process which is running on another CPU to enter
814 * kernel-mode, without any delay. (to get signals handled.)
816 void kick_process(task_t *p)
822 if ((cpu != smp_processor_id()) && task_curr(p))
823 smp_send_reschedule(cpu);
827 EXPORT_SYMBOL_GPL(kick_process);
830 * Return a low guess at the load of a migration-source cpu.
832 * We want to under-estimate the load of migration sources, to
833 * balance conservatively.
835 static inline unsigned long source_load(int cpu)
837 runqueue_t *rq = cpu_rq(cpu);
838 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
840 return min(rq->cpu_load, load_now);
844 * Return a high guess at the load of a migration-target cpu
846 static inline unsigned long target_load(int cpu)
848 runqueue_t *rq = cpu_rq(cpu);
849 unsigned long load_now = rq->nr_running * SCHED_LOAD_SCALE;
851 return max(rq->cpu_load, load_now);
857 * wake_idle() is useful especially on SMT architectures to wake a
858 * task onto an idle sibling if we would otherwise wake it onto a
861 * Returns the CPU we should wake onto.
863 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
864 static int wake_idle(int cpu, task_t *p)
867 runqueue_t *rq = cpu_rq(cpu);
868 struct sched_domain *sd;
875 if (!(sd->flags & SD_WAKE_IDLE))
878 cpus_and(tmp, sd->span, cpu_online_map);
879 cpus_and(tmp, tmp, p->cpus_allowed);
881 for_each_cpu_mask(i, tmp) {
889 static inline int wake_idle(int cpu, task_t *p)
896 * try_to_wake_up - wake up a thread
897 * @p: the to-be-woken-up thread
898 * @state: the mask of task states that can be woken
899 * @sync: do a synchronous wakeup?
901 * Put it on the run-queue if it's not already there. The "current"
902 * thread is always on the run-queue (except when the actual
903 * re-schedule is in progress), and as such you're allowed to do
904 * the simpler "current->state = TASK_RUNNING" to mark yourself
905 * runnable without the overhead of this.
907 * returns failure only if the task is already active.
909 static int try_to_wake_up(task_t * p, unsigned int state, int sync)
911 int cpu, this_cpu, success = 0;
916 unsigned long load, this_load;
917 struct sched_domain *sd;
921 rq = task_rq_lock(p, &flags);
922 old_state = p->state;
923 if (!(old_state & state))
930 this_cpu = smp_processor_id();
933 if (unlikely(task_running(rq, p)))
938 if (cpu == this_cpu || unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
941 load = source_load(cpu);
942 this_load = target_load(this_cpu);
945 * If sync wakeup then subtract the (maximum possible) effect of
946 * the currently running task from the load of the current CPU:
949 this_load -= SCHED_LOAD_SCALE;
951 /* Don't pull the task off an idle CPU to a busy one */
952 if (load < SCHED_LOAD_SCALE/2 && this_load > SCHED_LOAD_SCALE/2)
955 new_cpu = this_cpu; /* Wake to this CPU if we can */
958 * Scan domains for affine wakeup and passive balancing
961 for_each_domain(this_cpu, sd) {
962 unsigned int imbalance;
964 * Start passive balancing when half the imbalance_pct
967 imbalance = sd->imbalance_pct + (sd->imbalance_pct - 100) / 2;
969 if ( ((sd->flags & SD_WAKE_AFFINE) &&
970 !task_hot(p, rq->timestamp_last_tick, sd))
971 || ((sd->flags & SD_WAKE_BALANCE) &&
972 imbalance*this_load <= 100*load) ) {
974 * Now sd has SD_WAKE_AFFINE and p is cache cold in sd
975 * or sd has SD_WAKE_BALANCE and there is an imbalance
977 if (cpu_isset(cpu, sd->span))
982 new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
984 new_cpu = wake_idle(new_cpu, p);
985 if (new_cpu != cpu && cpu_isset(new_cpu, p->cpus_allowed)) {
986 set_task_cpu(p, new_cpu);
987 task_rq_unlock(rq, &flags);
988 /* might preempt at this point */
989 rq = task_rq_lock(p, &flags);
990 old_state = p->state;
991 if (!(old_state & state))
996 this_cpu = smp_processor_id();
1001 #endif /* CONFIG_SMP */
1002 if (old_state == TASK_UNINTERRUPTIBLE) {
1003 rq->nr_uninterruptible--;
1005 * Tasks on involuntary sleep don't earn
1006 * sleep_avg beyond just interactive state.
1012 * Sync wakeups (i.e. those types of wakeups where the waker
1013 * has indicated that it will leave the CPU in short order)
1014 * don't trigger a preemption, if the woken up task will run on
1015 * this cpu. (in this case the 'I will reschedule' promise of
1016 * the waker guarantees that the freshly woken up task is going
1017 * to be considered on this CPU.)
1019 activate_task(p, rq, cpu == this_cpu);
1020 if (!sync || cpu != this_cpu) {
1021 if (TASK_PREEMPTS_CURR(p, rq))
1022 resched_task(rq->curr);
1027 p->state = TASK_RUNNING;
1029 task_rq_unlock(rq, &flags);
1034 int fastcall wake_up_process(task_t * p)
1036 return try_to_wake_up(p, TASK_STOPPED |
1037 TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE, 0);
1040 EXPORT_SYMBOL(wake_up_process);
1042 int fastcall wake_up_state(task_t *p, unsigned int state)
1044 return try_to_wake_up(p, state, 0);
1048 * Perform scheduler related setup for a newly forked process p.
1049 * p is forked by current.
1051 void fastcall sched_fork(task_t *p)
1054 * We mark the process as running here, but have not actually
1055 * inserted it onto the runqueue yet. This guarantees that
1056 * nobody will actually run it, and a signal or other external
1057 * event cannot wake it up and insert it on the runqueue either.
1059 p->state = TASK_RUNNING;
1060 INIT_LIST_HEAD(&p->run_list);
1062 spin_lock_init(&p->switch_lock);
1063 #ifdef CONFIG_PREEMPT
1065 * During context-switch we hold precisely one spinlock, which
1066 * schedule_tail drops. (in the common case it's this_rq()->lock,
1067 * but it also can be p->switch_lock.) So we compensate with a count
1068 * of 1. Also, we want to start with kernel preemption disabled.
1070 p->thread_info->preempt_count = 1;
1073 * Share the timeslice between parent and child, thus the
1074 * total amount of pending timeslices in the system doesn't change,
1075 * resulting in more scheduling fairness.
1077 local_irq_disable();
1078 p->time_slice = (current->time_slice + 1) >> 1;
1080 * The remainder of the first timeslice might be recovered by
1081 * the parent if the child exits early enough.
1083 p->first_time_slice = 1;
1084 current->time_slice >>= 1;
1085 p->timestamp = sched_clock();
1086 if (!current->time_slice) {
1088 * This case is rare, it happens when the parent has only
1089 * a single jiffy left from its timeslice. Taking the
1090 * runqueue lock is not a problem.
1092 current->time_slice = 1;
1094 scheduler_tick(0, 0);
1102 * wake_up_forked_process - wake up a freshly forked process.
1104 * This function will do some initial scheduler statistics housekeeping
1105 * that must be done for every newly created process.
1107 void fastcall wake_up_forked_process(task_t * p)
1109 unsigned long flags;
1110 runqueue_t *rq = task_rq_lock(current, &flags);
1112 BUG_ON(p->state != TASK_RUNNING);
1115 * We decrease the sleep average of forking parents
1116 * and children as well, to keep max-interactive tasks
1117 * from forking tasks that are max-interactive.
1119 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1120 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1122 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1123 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1125 p->interactive_credit = 0;
1127 p->prio = effective_prio(p);
1128 set_task_cpu(p, smp_processor_id());
1130 if (unlikely(!current->array))
1131 __activate_task(p, rq);
1133 p->prio = current->prio;
1134 list_add_tail(&p->run_list, ¤t->run_list);
1135 p->array = current->array;
1136 p->array->nr_active++;
1140 task_rq_unlock(rq, &flags);
1144 * Potentially available exiting-child timeslices are
1145 * retrieved here - this way the parent does not get
1146 * penalized for creating too many threads.
1148 * (this cannot be used to 'generate' timeslices
1149 * artificially, because any timeslice recovered here
1150 * was given away by the parent in the first place.)
1152 void fastcall sched_exit(task_t * p)
1154 unsigned long flags;
1157 local_irq_save(flags);
1158 if (p->first_time_slice) {
1159 p->parent->time_slice += p->time_slice;
1160 if (unlikely(p->parent->time_slice > MAX_TIMESLICE))
1161 p->parent->time_slice = MAX_TIMESLICE;
1163 local_irq_restore(flags);
1165 * If the child was a (relative-) CPU hog then decrease
1166 * the sleep_avg of the parent as well.
1168 rq = task_rq_lock(p->parent, &flags);
1169 if (p->sleep_avg < p->parent->sleep_avg)
1170 p->parent->sleep_avg = p->parent->sleep_avg /
1171 (EXIT_WEIGHT + 1) * EXIT_WEIGHT + p->sleep_avg /
1173 task_rq_unlock(rq, &flags);
1177 * finish_task_switch - clean up after a task-switch
1178 * @prev: the thread we just switched away from.
1180 * We enter this with the runqueue still locked, and finish_arch_switch()
1181 * will unlock it along with doing any other architecture-specific cleanup
1184 * Note that we may have delayed dropping an mm in context_switch(). If
1185 * so, we finish that here outside of the runqueue lock. (Doing it
1186 * with the lock held can cause deadlocks; see schedule() for
1189 static void finish_task_switch(task_t *prev)
1191 runqueue_t *rq = this_rq();
1192 struct mm_struct *mm = rq->prev_mm;
1193 unsigned long prev_task_flags;
1198 * A task struct has one reference for the use as "current".
1199 * If a task dies, then it sets TASK_ZOMBIE in tsk->state and calls
1200 * schedule one last time. The schedule call will never return,
1201 * and the scheduled task must drop that reference.
1202 * The test for TASK_ZOMBIE must occur while the runqueue locks are
1203 * still held, otherwise prev could be scheduled on another cpu, die
1204 * there before we look at prev->state, and then the reference would
1206 * Manfred Spraul <manfred@colorfullife.com>
1208 prev_task_flags = prev->flags;
1209 finish_arch_switch(rq, prev);
1212 if (unlikely(prev_task_flags & PF_DEAD))
1213 put_task_struct(prev);
1217 * schedule_tail - first thing a freshly forked thread must call.
1218 * @prev: the thread we just switched away from.
1220 asmlinkage void schedule_tail(task_t *prev)
1222 finish_task_switch(prev);
1224 if (current->set_child_tid)
1225 put_user(current->pid, current->set_child_tid);
1229 * context_switch - switch to the new MM and the new
1230 * thread's register state.
1233 task_t * context_switch(runqueue_t *rq, task_t *prev, task_t *next)
1235 struct mm_struct *mm = next->mm;
1236 struct mm_struct *oldmm = prev->active_mm;
1238 if (unlikely(!mm)) {
1239 next->active_mm = oldmm;
1240 atomic_inc(&oldmm->mm_count);
1241 enter_lazy_tlb(oldmm, next);
1243 switch_mm(oldmm, mm, next);
1245 if (unlikely(!prev->mm)) {
1246 prev->active_mm = NULL;
1247 WARN_ON(rq->prev_mm);
1248 rq->prev_mm = oldmm;
1251 /* Here we just switch the register state and the stack. */
1252 switch_to(prev, next, prev);
1258 * nr_running, nr_uninterruptible and nr_context_switches:
1260 * externally visible scheduler statistics: current number of runnable
1261 * threads, current number of uninterruptible-sleeping threads, total
1262 * number of context switches performed since bootup.
1264 unsigned long nr_running(void)
1266 unsigned long i, sum = 0;
1269 sum += cpu_rq(i)->nr_running;
1274 unsigned long nr_uninterruptible(void)
1276 unsigned long i, sum = 0;
1278 for_each_online_cpu(i)
1279 sum += cpu_rq(i)->nr_uninterruptible;
1284 unsigned long long nr_context_switches(void)
1286 unsigned long long i, sum = 0;
1288 for_each_online_cpu(i)
1289 sum += cpu_rq(i)->nr_switches;
1294 unsigned long nr_iowait(void)
1296 unsigned long i, sum = 0;
1298 for_each_online_cpu(i)
1299 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1305 * double_rq_lock - safely lock two runqueues
1307 * Note this does not disable interrupts like task_rq_lock,
1308 * you need to do so manually before calling.
1310 static void double_rq_lock(runqueue_t *rq1, runqueue_t *rq2)
1313 spin_lock(&rq1->lock);
1316 spin_lock(&rq1->lock);
1317 spin_lock(&rq2->lock);
1319 spin_lock(&rq2->lock);
1320 spin_lock(&rq1->lock);
1326 * double_rq_unlock - safely unlock two runqueues
1328 * Note this does not restore interrupts like task_rq_unlock,
1329 * you need to do so manually after calling.
1331 static void double_rq_unlock(runqueue_t *rq1, runqueue_t *rq2)
1333 spin_unlock(&rq1->lock);
1335 spin_unlock(&rq2->lock);
1338 unsigned long long nr_preempt(void)
1340 unsigned long long i, sum = 0;
1342 for_each_online_cpu(i)
1343 sum += cpu_rq(i)->nr_preempt;
1358 * find_idlest_cpu - find the least busy runqueue.
1360 static int find_idlest_cpu(struct task_struct *p, int this_cpu,
1361 struct sched_domain *sd)
1363 unsigned long load, min_load, this_load;
1368 min_load = ULONG_MAX;
1370 cpus_and(mask, sd->span, cpu_online_map);
1371 cpus_and(mask, mask, p->cpus_allowed);
1373 for_each_cpu_mask(i, mask) {
1374 load = target_load(i);
1376 if (load < min_load) {
1380 /* break out early on an idle CPU: */
1386 /* add +1 to account for the new task */
1387 this_load = source_load(this_cpu) + SCHED_LOAD_SCALE;
1390 * Would with the addition of the new task to the
1391 * current CPU there be an imbalance between this
1392 * CPU and the idlest CPU?
1394 * Use half of the balancing threshold - new-context is
1395 * a good opportunity to balance.
1397 if (min_load*(100 + (sd->imbalance_pct-100)/2) < this_load*100)
1404 * wake_up_forked_thread - wake up a freshly forked thread.
1406 * This function will do some initial scheduler statistics housekeeping
1407 * that must be done for every newly created context, and it also does
1408 * runqueue balancing.
1410 void fastcall wake_up_forked_thread(task_t * p)
1412 unsigned long flags;
1413 int this_cpu = get_cpu(), cpu;
1414 struct sched_domain *tmp, *sd = NULL;
1415 runqueue_t *this_rq = cpu_rq(this_cpu), *rq;
1418 * Find the largest domain that this CPU is part of that
1419 * is willing to balance on clone:
1421 for_each_domain(this_cpu, tmp)
1422 if (tmp->flags & SD_BALANCE_CLONE)
1425 cpu = find_idlest_cpu(p, this_cpu, sd);
1429 local_irq_save(flags);
1432 double_rq_lock(this_rq, rq);
1434 BUG_ON(p->state != TASK_RUNNING);
1437 * We did find_idlest_cpu() unlocked, so in theory
1438 * the mask could have changed - just dont migrate
1441 if (unlikely(!cpu_isset(cpu, p->cpus_allowed))) {
1443 double_rq_unlock(this_rq, rq);
1447 * We decrease the sleep average of forking parents
1448 * and children as well, to keep max-interactive tasks
1449 * from forking tasks that are max-interactive.
1451 current->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(current) *
1452 PARENT_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1454 p->sleep_avg = JIFFIES_TO_NS(CURRENT_BONUS(p) *
1455 CHILD_PENALTY / 100 * MAX_SLEEP_AVG / MAX_BONUS);
1457 p->interactive_credit = 0;
1459 p->prio = effective_prio(p);
1460 set_task_cpu(p, cpu);
1462 if (cpu == this_cpu) {
1463 if (unlikely(!current->array))
1464 __activate_task(p, rq);
1466 p->prio = current->prio;
1467 list_add_tail(&p->run_list, ¤t->run_list);
1468 p->array = current->array;
1469 p->array->nr_active++;
1474 /* Not the local CPU - must adjust timestamp */
1475 p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
1476 + rq->timestamp_last_tick;
1477 __activate_task(p, rq);
1478 if (TASK_PREEMPTS_CURR(p, rq))
1479 resched_task(rq->curr);
1482 double_rq_unlock(this_rq, rq);
1483 local_irq_restore(flags);
1488 * If dest_cpu is allowed for this process, migrate the task to it.
1489 * This is accomplished by forcing the cpu_allowed mask to only
1490 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
1491 * the cpu_allowed mask is restored.
1493 static void sched_migrate_task(task_t *p, int dest_cpu)
1495 migration_req_t req;
1497 unsigned long flags;
1499 rq = task_rq_lock(p, &flags);
1500 if (!cpu_isset(dest_cpu, p->cpus_allowed)
1501 || unlikely(cpu_is_offline(dest_cpu)))
1504 /* force the process onto the specified CPU */
1505 if (migrate_task(p, dest_cpu, &req)) {
1506 /* Need to wait for migration thread (might exit: take ref). */
1507 struct task_struct *mt = rq->migration_thread;
1508 get_task_struct(mt);
1509 task_rq_unlock(rq, &flags);
1510 wake_up_process(mt);
1511 put_task_struct(mt);
1512 wait_for_completion(&req.done);
1516 task_rq_unlock(rq, &flags);
1520 * sched_balance_exec(): find the highest-level, exec-balance-capable
1521 * domain and try to migrate the task to the least loaded CPU.
1523 * execve() is a valuable balancing opportunity, because at this point
1524 * the task has the smallest effective memory and cache footprint.
1526 void sched_balance_exec(void)
1528 struct sched_domain *tmp, *sd = NULL;
1529 int new_cpu, this_cpu = get_cpu();
1531 /* Prefer the current CPU if there's only this task running */
1532 if (this_rq()->nr_running <= 1)
1535 for_each_domain(this_cpu, tmp)
1536 if (tmp->flags & SD_BALANCE_EXEC)
1540 new_cpu = find_idlest_cpu(current, this_cpu, sd);
1541 if (new_cpu != this_cpu) {
1543 sched_migrate_task(current, new_cpu);
1552 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1554 static void double_lock_balance(runqueue_t *this_rq, runqueue_t *busiest)
1556 if (unlikely(!spin_trylock(&busiest->lock))) {
1557 if (busiest < this_rq) {
1558 spin_unlock(&this_rq->lock);
1559 spin_lock(&busiest->lock);
1560 spin_lock(&this_rq->lock);
1562 spin_lock(&busiest->lock);
1567 * pull_task - move a task from a remote runqueue to the local runqueue.
1568 * Both runqueues must be locked.
1571 void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
1572 runqueue_t *this_rq, prio_array_t *this_array, int this_cpu)
1574 dequeue_task(p, src_array);
1575 src_rq->nr_running--;
1576 rq_load_dec(src_rq,p);
1578 set_task_cpu(p, this_cpu);
1579 this_rq->nr_running++;
1580 rq_load_inc(this_rq,p);
1581 enqueue_task(p, this_array);
1583 p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
1584 + this_rq->timestamp_last_tick;
1586 * Note that idle threads have a prio of MAX_PRIO, for this test
1587 * to be always true for them.
1589 if (TASK_PREEMPTS_CURR(p, this_rq))
1590 resched_task(this_rq->curr);
1594 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
1597 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
1598 struct sched_domain *sd, enum idle_type idle)
1601 * We do not migrate tasks that are:
1602 * 1) running (obviously), or
1603 * 2) cannot be migrated to this CPU due to cpus_allowed, or
1604 * 3) are cache-hot on their current CPU.
1606 if (task_running(rq, p))
1608 if (!cpu_isset(this_cpu, p->cpus_allowed))
1611 /* Aggressive migration if we've failed balancing */
1612 if (idle == NEWLY_IDLE ||
1613 sd->nr_balance_failed < sd->cache_nice_tries) {
1614 if (task_hot(p, rq->timestamp_last_tick, sd))
1621 #ifdef CONFIG_CKRM_CPU_SCHEDULE
1623 struct ckrm_cpu_class *find_unbalanced_class(int busiest_cpu, int this_cpu, unsigned long *cls_imbalance)
1625 struct ckrm_cpu_class *most_unbalanced_class = NULL;
1626 struct ckrm_cpu_class *clsptr;
1627 int max_unbalance = 0;
1629 list_for_each_entry(clsptr,&active_cpu_classes,links) {
1630 struct ckrm_local_runqueue *this_lrq = get_ckrm_local_runqueue(clsptr,this_cpu);
1631 struct ckrm_local_runqueue *busiest_lrq = get_ckrm_local_runqueue(clsptr,busiest_cpu);
1632 int unbalance_degree;
1634 unbalance_degree = (local_queue_nr_running(busiest_lrq) - local_queue_nr_running(this_lrq)) * cpu_class_weight(clsptr);
1635 if (unbalance_degree >= *cls_imbalance)
1636 continue; // already looked at this class
1638 if (unbalance_degree > max_unbalance) {
1639 max_unbalance = unbalance_degree;
1640 most_unbalanced_class = clsptr;
1643 *cls_imbalance = max_unbalance;
1644 return most_unbalanced_class;
1649 * find_busiest_queue - find the busiest runqueue among the cpus in cpumask.
1651 static int find_busiest_cpu(runqueue_t *this_rq, int this_cpu, int idle,
1654 int cpu_load, load, max_load, i, busiest_cpu;
1655 runqueue_t *busiest, *rq_src;
1658 /*Hubertus ... the concept of nr_running is replace with cpu_load */
1659 cpu_load = this_rq->ckrm_cpu_load;
1665 for_each_online_cpu(i) {
1667 load = rq_src->ckrm_cpu_load;
1669 if ((load > max_load) && (rq_src != this_rq)) {
1676 if (likely(!busiest))
1679 *imbalance = max_load - cpu_load;
1681 /* It needs an at least ~25% imbalance to trigger balancing. */
1682 if (!idle && ((*imbalance)*4 < max_load)) {
1687 double_lock_balance(this_rq, busiest);
1689 * Make sure nothing changed since we checked the
1692 if (busiest->ckrm_cpu_load <= cpu_load) {
1693 spin_unlock(&busiest->lock);
1697 return (busiest ? busiest_cpu : -1);
1700 static int load_balance(int this_cpu, runqueue_t *this_rq,
1701 struct sched_domain *sd, enum idle_type idle)
1705 runqueue_t *busiest;
1706 prio_array_t *array;
1707 struct list_head *head, *curr;
1709 struct ckrm_local_runqueue * busiest_local_queue;
1710 struct ckrm_cpu_class *clsptr;
1712 unsigned long cls_imbalance; // so we can retry other classes
1714 // need to update global CVT based on local accumulated CVTs
1715 read_lock(&class_list_lock);
1716 busiest_cpu = find_busiest_cpu(this_rq, this_cpu, idle, &imbalance);
1717 if (busiest_cpu == -1)
1720 busiest = cpu_rq(busiest_cpu);
1723 * We only want to steal a number of tasks equal to 1/2 the imbalance,
1724 * otherwise we'll just shift the imbalance to the new queue:
1728 /* now find class on that runqueue with largest inbalance */
1729 cls_imbalance = 0xFFFFFFFF;
1732 clsptr = find_unbalanced_class(busiest_cpu, this_cpu, &cls_imbalance);
1736 busiest_local_queue = get_ckrm_local_runqueue(clsptr,busiest_cpu);
1737 weight = cpu_class_weight(clsptr);
1740 * We first consider expired tasks. Those will likely not be
1741 * executed in the near future, and they are most likely to
1742 * be cache-cold, thus switching CPUs has the least effect
1745 if (busiest_local_queue->expired->nr_active)
1746 array = busiest_local_queue->expired;
1748 array = busiest_local_queue->active;
1751 /* Start searching at priority 0: */
1755 idx = sched_find_first_bit(array->bitmap);
1757 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1758 if (idx >= MAX_PRIO) {
1759 if (array == busiest_local_queue->expired && busiest_local_queue->active->nr_active) {
1760 array = busiest_local_queue->active;
1763 goto retry_other_class;
1766 head = array->queue + idx;
1769 tmp = list_entry(curr, task_t, run_list);
1773 if (!can_migrate_task(tmp, busiest, this_cpu, sd,idle)) {
1779 pull_task(busiest, array, tmp, this_rq, rq_active(tmp,this_rq),this_cpu);
1781 * tmp BUG FIX: hzheng
1782 * load balancing can make the busiest local queue empty
1783 * thus it should be removed from bpt
1785 if (! local_queue_nr_running(busiest_local_queue)) {
1786 classqueue_dequeue(busiest_local_queue->classqueue,&busiest_local_queue->classqueue_linkobj);
1787 cpu_demand_event(get_rq_local_stat(busiest_local_queue,busiest_cpu),CPU_DEMAND_DEQUEUE,0);
1790 imbalance -= weight;
1791 if (!idle && (imbalance>0)) {
1798 spin_unlock(&busiest->lock);
1800 read_unlock(&class_list_lock);
1805 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
1808 #else /* CONFIG_CKRM_CPU_SCHEDULE */
1810 * move_tasks tries to move up to max_nr_move tasks from busiest to this_rq,
1811 * as part of a balancing operation within "domain". Returns the number of
1814 * Called with both runqueues locked.
1816 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
1817 unsigned long max_nr_move, struct sched_domain *sd,
1818 enum idle_type idle)
1820 prio_array_t *array, *dst_array;
1821 struct list_head *head, *curr;
1822 int idx, pulled = 0;
1825 if (max_nr_move <= 0 || busiest->nr_running <= 1)
1829 * We first consider expired tasks. Those will likely not be
1830 * executed in the near future, and they are most likely to
1831 * be cache-cold, thus switching CPUs has the least effect
1834 if (busiest->expired->nr_active) {
1835 array = busiest->expired;
1836 dst_array = this_rq->expired;
1838 array = busiest->active;
1839 dst_array = this_rq->active;
1843 /* Start searching at priority 0: */
1847 idx = sched_find_first_bit(array->bitmap);
1849 idx = find_next_bit(array->bitmap, MAX_PRIO, idx);
1850 if (idx >= MAX_PRIO) {
1851 if (array == busiest->expired && busiest->active->nr_active) {
1852 array = busiest->active;
1853 dst_array = this_rq->active;
1859 head = array->queue + idx;
1862 tmp = list_entry(curr, task_t, run_list);
1866 if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle)) {
1872 pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
1875 /* We only want to steal up to the prescribed number of tasks. */
1876 if (pulled < max_nr_move) {
1887 * find_busiest_group finds and returns the busiest CPU group within the
1888 * domain. It calculates and returns the number of tasks which should be
1889 * moved to restore balance via the imbalance parameter.
1891 static struct sched_group *
1892 find_busiest_group(struct sched_domain *sd, int this_cpu,
1893 unsigned long *imbalance, enum idle_type idle)
1895 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
1896 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
1898 max_load = this_load = total_load = total_pwr = 0;
1906 local_group = cpu_isset(this_cpu, group->cpumask);
1908 /* Tally up the load of all CPUs in the group */
1910 cpus_and(tmp, group->cpumask, cpu_online_map);
1911 if (unlikely(cpus_empty(tmp)))
1914 for_each_cpu_mask(i, tmp) {
1915 /* Bias balancing toward cpus of our domain */
1917 load = target_load(i);
1919 load = source_load(i);
1928 total_load += avg_load;
1929 total_pwr += group->cpu_power;
1931 /* Adjust by relative CPU power of the group */
1932 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
1935 this_load = avg_load;
1938 } else if (avg_load > max_load) {
1939 max_load = avg_load;
1943 group = group->next;
1944 } while (group != sd->groups);
1946 if (!busiest || this_load >= max_load)
1949 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
1951 if (this_load >= avg_load ||
1952 100*max_load <= sd->imbalance_pct*this_load)
1956 * We're trying to get all the cpus to the average_load, so we don't
1957 * want to push ourselves above the average load, nor do we wish to
1958 * reduce the max loaded cpu below the average load, as either of these
1959 * actions would just result in more rebalancing later, and ping-pong
1960 * tasks around. Thus we look for the minimum possible imbalance.
1961 * Negative imbalances (*we* are more loaded than anyone else) will
1962 * be counted as no imbalance for these purposes -- we can't fix that
1963 * by pulling tasks to us. Be careful of negative numbers as they'll
1964 * appear as very large values with unsigned longs.
1966 *imbalance = min(max_load - avg_load, avg_load - this_load);
1968 /* How much load to actually move to equalise the imbalance */
1969 *imbalance = (*imbalance * min(busiest->cpu_power, this->cpu_power))
1972 if (*imbalance < SCHED_LOAD_SCALE - 1) {
1973 unsigned long pwr_now = 0, pwr_move = 0;
1976 if (max_load - this_load >= SCHED_LOAD_SCALE*2) {
1982 * OK, we don't have enough imbalance to justify moving tasks,
1983 * however we may be able to increase total CPU power used by
1987 pwr_now += busiest->cpu_power*min(SCHED_LOAD_SCALE, max_load);
1988 pwr_now += this->cpu_power*min(SCHED_LOAD_SCALE, this_load);
1989 pwr_now /= SCHED_LOAD_SCALE;
1991 /* Amount of load we'd subtract */
1992 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/busiest->cpu_power;
1994 pwr_move += busiest->cpu_power*min(SCHED_LOAD_SCALE,
1997 /* Amount of load we'd add */
1998 tmp = SCHED_LOAD_SCALE*SCHED_LOAD_SCALE/this->cpu_power;
2001 pwr_move += this->cpu_power*min(SCHED_LOAD_SCALE, this_load + tmp);
2002 pwr_move /= SCHED_LOAD_SCALE;
2004 /* Move if we gain another 8th of a CPU worth of throughput */
2005 if (pwr_move < pwr_now + SCHED_LOAD_SCALE / 8)
2012 /* Get rid of the scaling factor, rounding down as we divide */
2013 *imbalance = (*imbalance + 1) / SCHED_LOAD_SCALE;
2018 if (busiest && (idle == NEWLY_IDLE ||
2019 (idle == IDLE && max_load > SCHED_LOAD_SCALE)) ) {
2029 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2031 static runqueue_t *find_busiest_queue(struct sched_group *group)
2034 unsigned long load, max_load = 0;
2035 runqueue_t *busiest = NULL;
2038 cpus_and(tmp, group->cpumask, cpu_online_map);
2039 for_each_cpu_mask(i, tmp) {
2040 load = source_load(i);
2042 if (load > max_load) {
2044 busiest = cpu_rq(i);
2052 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2053 * tasks if there is an imbalance.
2055 * Called with this_rq unlocked.
2057 static int load_balance(int this_cpu, runqueue_t *this_rq,
2058 struct sched_domain *sd, enum idle_type idle)
2060 struct sched_group *group;
2061 runqueue_t *busiest;
2062 unsigned long imbalance;
2065 spin_lock(&this_rq->lock);
2067 group = find_busiest_group(sd, this_cpu, &imbalance, idle);
2071 busiest = find_busiest_queue(group);
2075 * This should be "impossible", but since load
2076 * balancing is inherently racy and statistical,
2077 * it could happen in theory.
2079 if (unlikely(busiest == this_rq)) {
2085 if (busiest->nr_running > 1) {
2087 * Attempt to move tasks. If find_busiest_group has found
2088 * an imbalance but busiest->nr_running <= 1, the group is
2089 * still unbalanced. nr_moved simply stays zero, so it is
2090 * correctly treated as an imbalance.
2092 double_lock_balance(this_rq, busiest);
2093 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2094 imbalance, sd, idle);
2095 spin_unlock(&busiest->lock);
2097 spin_unlock(&this_rq->lock);
2100 sd->nr_balance_failed++;
2102 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
2105 spin_lock(&busiest->lock);
2106 if (!busiest->active_balance) {
2107 busiest->active_balance = 1;
2108 busiest->push_cpu = this_cpu;
2111 spin_unlock(&busiest->lock);
2113 wake_up_process(busiest->migration_thread);
2116 * We've kicked active balancing, reset the failure
2119 sd->nr_balance_failed = sd->cache_nice_tries;
2122 sd->nr_balance_failed = 0;
2124 /* We were unbalanced, so reset the balancing interval */
2125 sd->balance_interval = sd->min_interval;
2130 spin_unlock(&this_rq->lock);
2132 /* tune up the balancing interval */
2133 if (sd->balance_interval < sd->max_interval)
2134 sd->balance_interval *= 2;
2140 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2141 * tasks if there is an imbalance.
2143 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2144 * this_rq is locked.
2146 static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
2147 struct sched_domain *sd)
2149 struct sched_group *group;
2150 runqueue_t *busiest = NULL;
2151 unsigned long imbalance;
2154 group = find_busiest_group(sd, this_cpu, &imbalance, NEWLY_IDLE);
2158 busiest = find_busiest_queue(group);
2159 if (!busiest || busiest == this_rq)
2162 /* Attempt to move tasks */
2163 double_lock_balance(this_rq, busiest);
2165 nr_moved = move_tasks(this_rq, this_cpu, busiest,
2166 imbalance, sd, NEWLY_IDLE);
2168 spin_unlock(&busiest->lock);
2175 * idle_balance is called by schedule() if this_cpu is about to become
2176 * idle. Attempts to pull tasks from other CPUs.
2178 static inline void idle_balance(int this_cpu, runqueue_t *this_rq)
2180 struct sched_domain *sd;
2182 for_each_domain(this_cpu, sd) {
2183 if (sd->flags & SD_BALANCE_NEWIDLE) {
2184 if (load_balance_newidle(this_cpu, this_rq, sd)) {
2185 /* We've pulled tasks over so stop searching */
2193 * active_load_balance is run by migration threads. It pushes a running
2194 * task off the cpu. It can be required to correctly have at least 1 task
2195 * running on each physical CPU where possible, and not have a physical /
2196 * logical imbalance.
2198 * Called with busiest locked.
2200 static void active_load_balance(runqueue_t *busiest, int busiest_cpu)
2202 struct sched_domain *sd;
2203 struct sched_group *group, *busy_group;
2206 if (busiest->nr_running <= 1)
2209 for_each_domain(busiest_cpu, sd)
2210 if (cpu_isset(busiest->push_cpu, sd->span))
2218 while (!cpu_isset(busiest_cpu, group->cpumask))
2219 group = group->next;
2228 if (group == busy_group)
2231 cpus_and(tmp, group->cpumask, cpu_online_map);
2232 if (!cpus_weight(tmp))
2235 for_each_cpu_mask(i, tmp) {
2241 rq = cpu_rq(push_cpu);
2244 * This condition is "impossible", but since load
2245 * balancing is inherently a bit racy and statistical,
2246 * it can trigger.. Reported by Bjorn Helgaas on a
2249 if (unlikely(busiest == rq))
2251 double_lock_balance(busiest, rq);
2252 move_tasks(rq, push_cpu, busiest, 1, sd, IDLE);
2253 spin_unlock(&rq->lock);
2255 group = group->next;
2256 } while (group != sd->groups);
2258 #endif /* CONFIG_CKRM_CPU_SCHEDULE*/
2261 * rebalance_tick will get called every timer tick, on every CPU.
2263 * It checks each scheduling domain to see if it is due to be balanced,
2264 * and initiates a balancing operation if so.
2266 * Balancing parameters are set up in arch_init_sched_domains.
2269 /* Don't have all balancing operations going off at once */
2270 #define CPU_OFFSET(cpu) (HZ * cpu / NR_CPUS)
2272 static void rebalance_tick(int this_cpu, runqueue_t *this_rq,
2273 enum idle_type idle)
2275 unsigned long old_load, this_load;
2276 unsigned long j = jiffies + CPU_OFFSET(this_cpu);
2277 struct sched_domain *sd;
2279 ckrm_rebalance_tick(j,this_cpu);
2281 /* Update our load */
2282 old_load = this_rq->cpu_load;
2283 this_load = this_rq->nr_running * SCHED_LOAD_SCALE;
2285 * Round up the averaging division if load is increasing. This
2286 * prevents us from getting stuck on 9 if the load is 10, for
2289 if (this_load > old_load)
2291 this_rq->cpu_load = (old_load + this_load) / 2;
2293 for_each_domain(this_cpu, sd) {
2294 unsigned long interval = sd->balance_interval;
2297 interval *= sd->busy_factor;
2299 /* scale ms to jiffies */
2300 interval = msecs_to_jiffies(interval);
2301 if (unlikely(!interval))
2304 if (j - sd->last_balance >= interval) {
2305 if (load_balance(this_cpu, this_rq, sd, idle)) {
2306 /* We've pulled tasks over so no longer idle */
2309 sd->last_balance += interval;
2315 * on UP we do not need to balance between CPUs:
2317 static inline void rebalance_tick(int cpu, runqueue_t *rq, enum idle_type idle)
2319 ckrm_rebalance_tick(jiffies,cpu);
2322 static inline void idle_balance(int cpu, runqueue_t *rq)
2327 static inline int wake_priority_sleeper(runqueue_t *rq)
2329 #ifdef CONFIG_SCHED_SMT
2331 * If an SMT sibling task has been put to sleep for priority
2332 * reasons reschedule the idle task to see if it can now run.
2334 if (rq->nr_running) {
2335 resched_task(rq->idle);
2342 DEFINE_PER_CPU(struct kernel_stat, kstat) = { { 0 } };
2344 EXPORT_PER_CPU_SYMBOL(kstat);
2347 * We place interactive tasks back into the active array, if possible.
2349 * To guarantee that this does not starve expired tasks we ignore the
2350 * interactivity of a task if the first expired task had to wait more
2351 * than a 'reasonable' amount of time. This deadline timeout is
2352 * load-dependent, as the frequency of array switched decreases with
2353 * increasing number of running tasks. We also ignore the interactivity
2354 * if a better static_prio task has expired:
2357 #ifndef CONFIG_CKRM_CPU_SCHEDULE
2358 #define EXPIRED_STARVING(rq) \
2359 ((STARVATION_LIMIT && ((rq)->expired_timestamp && \
2360 (jiffies - (rq)->expired_timestamp >= \
2361 STARVATION_LIMIT * ((rq)->nr_running) + 1))) || \
2362 ((rq)->curr->static_prio > (rq)->best_expired_prio))
2364 #define EXPIRED_STARVING(rq) \
2365 (STARVATION_LIMIT && ((rq)->expired_timestamp && \
2366 (jiffies - (rq)->expired_timestamp >= \
2367 STARVATION_LIMIT * (local_queue_nr_running(rq)) + 1)))
2371 * This function gets called by the timer code, with HZ frequency.
2372 * We call it with interrupts disabled.
2374 * It also gets called by the fork code, when changing the parent's
2377 void scheduler_tick(int user_ticks, int sys_ticks)
2379 int cpu = smp_processor_id();
2380 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2381 runqueue_t *rq = this_rq();
2382 task_t *p = current;
2384 rq->timestamp_last_tick = sched_clock();
2386 if (rcu_pending(cpu))
2387 rcu_check_callbacks(cpu, user_ticks);
2389 /* note: this timer irq context must be accounted for as well */
2390 if (hardirq_count() - HARDIRQ_OFFSET) {
2391 cpustat->irq += sys_ticks;
2393 } else if (softirq_count()) {
2394 cpustat->softirq += sys_ticks;
2398 if (p == rq->idle) {
2399 if (!--rq->idle_tokens && !list_empty(&rq->hold_queue))
2402 if (atomic_read(&rq->nr_iowait) > 0)
2403 cpustat->iowait += sys_ticks;
2405 cpustat->idle += sys_ticks;
2406 if (wake_priority_sleeper(rq))
2408 rebalance_tick(cpu, rq, IDLE);
2411 if (TASK_NICE(p) > 0)
2412 cpustat->nice += user_ticks;
2414 cpustat->user += user_ticks;
2415 cpustat->system += sys_ticks;
2417 /* Task might have expired already, but not scheduled off yet */
2418 if (p->array != rq_active(p,rq)) {
2419 set_tsk_need_resched(p);
2422 spin_lock(&rq->lock);
2424 * The task was running during this tick - update the
2425 * time slice counter. Note: we do not update a thread's
2426 * priority until it either goes to sleep or uses up its
2427 * timeslice. This makes it possible for interactive tasks
2428 * to use up their timeslices at their highest priority levels.
2430 if (unlikely(rt_task(p))) {
2432 * RR tasks need a special form of timeslice management.
2433 * FIFO tasks have no timeslices.
2435 if ((p->policy == SCHED_RR) && !--p->time_slice) {
2436 p->time_slice = task_timeslice(p);
2437 p->first_time_slice = 0;
2438 set_tsk_need_resched(p);
2440 /* put it at the end of the queue: */
2441 dequeue_task(p, rq_active(p,rq));
2442 enqueue_task(p, rq_active(p,rq));
2446 #warning MEF PLANETLAB: "if (vx_need_resched(p)) was if (!--p->time_slice) */"
2447 if (vx_need_resched(p)) {
2448 #ifdef CONFIG_CKRM_CPU_SCHEDULE
2449 /* Hubertus ... we can abstract this out */
2450 struct ckrm_local_runqueue* rq = get_task_class_queue(p);
2452 dequeue_task(p, rq->active);
2453 set_tsk_need_resched(p);
2454 p->prio = effective_prio(p);
2455 p->time_slice = task_timeslice(p);
2456 p->first_time_slice = 0;
2458 if (!rq->expired_timestamp)
2459 rq->expired_timestamp = jiffies;
2460 if (!TASK_INTERACTIVE(p) || EXPIRED_STARVING(rq)) {
2461 enqueue_task(p, rq->expired);
2462 if (p->static_prio < this_rq()->best_expired_prio)
2463 this_rq()->best_expired_prio = p->static_prio;
2465 enqueue_task(p, rq->active);
2468 * Prevent a too long timeslice allowing a task to monopolize
2469 * the CPU. We do this by splitting up the timeslice into
2472 * Note: this does not mean the task's timeslices expire or
2473 * get lost in any way, they just might be preempted by
2474 * another task of equal priority. (one with higher
2475 * priority would have preempted this task already.) We
2476 * requeue this task to the end of the list on this priority
2477 * level, which is in essence a round-robin of tasks with
2480 * This only applies to tasks in the interactive
2481 * delta range with at least TIMESLICE_GRANULARITY to requeue.
2483 if (TASK_INTERACTIVE(p) && !((task_timeslice(p) -
2484 p->time_slice) % TIMESLICE_GRANULARITY(p)) &&
2485 (p->time_slice >= TIMESLICE_GRANULARITY(p)) &&
2486 (p->array == rq_active(p,rq))) {
2488 dequeue_task(p, rq_active(p,rq));
2489 set_tsk_need_resched(p);
2490 p->prio = effective_prio(p);
2491 enqueue_task(p, rq_active(p,rq));
2495 spin_unlock(&rq->lock);
2497 rebalance_tick(cpu, rq, NOT_IDLE);
2500 #ifdef CONFIG_SCHED_SMT
2501 static inline void wake_sleeping_dependent(int cpu, runqueue_t *rq)
2504 struct sched_domain *sd = rq->sd;
2505 cpumask_t sibling_map;
2507 if (!(sd->flags & SD_SHARE_CPUPOWER))
2510 cpus_and(sibling_map, sd->span, cpu_online_map);
2511 for_each_cpu_mask(i, sibling_map) {
2520 * If an SMT sibling task is sleeping due to priority
2521 * reasons wake it up now.
2523 if (smt_rq->curr == smt_rq->idle && smt_rq->nr_running)
2524 resched_task(smt_rq->idle);
2528 static inline int dependent_sleeper(int cpu, runqueue_t *rq, task_t *p)
2530 struct sched_domain *sd = rq->sd;
2531 cpumask_t sibling_map;
2534 if (!(sd->flags & SD_SHARE_CPUPOWER))
2537 cpus_and(sibling_map, sd->span, cpu_online_map);
2538 for_each_cpu_mask(i, sibling_map) {
2546 smt_curr = smt_rq->curr;
2549 * If a user task with lower static priority than the
2550 * running task on the SMT sibling is trying to schedule,
2551 * delay it till there is proportionately less timeslice
2552 * left of the sibling task to prevent a lower priority
2553 * task from using an unfair proportion of the
2554 * physical cpu's resources. -ck
2556 if (((smt_curr->time_slice * (100 - sd->per_cpu_gain) / 100) >
2557 task_timeslice(p) || rt_task(smt_curr)) &&
2558 p->mm && smt_curr->mm && !rt_task(p))
2562 * Reschedule a lower priority task on the SMT sibling,
2563 * or wake it up if it has been put to sleep for priority
2566 if ((((p->time_slice * (100 - sd->per_cpu_gain) / 100) >
2567 task_timeslice(smt_curr) || rt_task(p)) &&
2568 smt_curr->mm && p->mm && !rt_task(smt_curr)) ||
2569 (smt_curr == smt_rq->idle && smt_rq->nr_running))
2570 resched_task(smt_curr);
2575 static inline void wake_sleeping_dependent(int cpu, runqueue_t *rq)
2579 static inline int dependent_sleeper(int cpu, runqueue_t *rq, task_t *p)
2586 * schedule() is the main scheduler function.
2588 asmlinkage void __sched schedule(void)
2591 task_t *prev, *next;
2593 prio_array_t *array;
2594 unsigned long long now;
2595 unsigned long run_time;
2597 #ifdef CONFIG_VSERVER_HARDCPU
2598 struct vx_info *vxi;
2602 //WARN_ON(system_state == SYSTEM_BOOTING);
2604 * Test if we are atomic. Since do_exit() needs to call into
2605 * schedule() atomically, we ignore that path for now.
2606 * Otherwise, whine if we are scheduling when we should not be.
2608 if (likely(!(current->state & (TASK_DEAD | TASK_ZOMBIE)))) {
2609 if (unlikely(in_atomic())) {
2610 printk(KERN_ERR "bad: scheduling while atomic!\n");
2620 release_kernel_lock(prev);
2621 now = sched_clock();
2622 if (likely(now - prev->timestamp < NS_MAX_SLEEP_AVG))
2623 run_time = now - prev->timestamp;
2625 run_time = NS_MAX_SLEEP_AVG;
2628 * Tasks with interactive credits get charged less run_time
2629 * at high sleep_avg to delay them losing their interactive
2632 if (HIGH_CREDIT(prev))
2633 run_time /= (CURRENT_BONUS(prev) ? : 1);
2635 spin_lock_irq(&rq->lock);
2638 * if entering off of a kernel preemption go straight
2639 * to picking the next task.
2641 switch_count = &prev->nivcsw;
2642 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2643 switch_count = &prev->nvcsw;
2644 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
2645 unlikely(signal_pending(prev))))
2646 prev->state = TASK_RUNNING;
2648 deactivate_task(prev, rq);
2651 cpu = smp_processor_id();
2652 #ifdef CONFIG_VSERVER_HARDCPU
2653 if (!list_empty(&rq->hold_queue)) {
2654 struct list_head *l, *n;
2658 list_for_each_safe(l, n, &rq->hold_queue) {
2659 next = list_entry(l, task_t, run_list);
2660 if (vxi == next->vx_info)
2663 vxi = next->vx_info;
2664 ret = vx_tokens_recalc(vxi);
2665 // tokens = vx_tokens_avail(next);
2668 list_del(&next->run_list);
2669 next->state &= ~TASK_ONHOLD;
2670 recalc_task_prio(next, now);
2671 __activate_task(next, rq);
2672 // printk("··· unhold %p\n", next);
2675 if ((ret < 0) && (maxidle < ret))
2679 rq->idle_tokens = -maxidle;
2683 if (unlikely(!rq->nr_running)) {
2684 idle_balance(cpu, rq);
2685 if (!rq->nr_running) {
2687 rq->expired_timestamp = 0;
2688 wake_sleeping_dependent(cpu, rq);
2693 next = rq_get_next_task(rq);
2694 if (next == rq->idle)
2697 if (dependent_sleeper(cpu, rq, next)) {
2702 #ifdef CONFIG_VSERVER_HARDCPU
2703 vxi = next->vx_info;
2704 if (vxi && __vx_flags(vxi->vx_flags,
2705 VXF_SCHED_PAUSE|VXF_SCHED_HARD, 0)) {
2706 int ret = vx_tokens_recalc(vxi);
2708 if (unlikely(ret <= 0)) {
2709 if (ret && (rq->idle_tokens > -ret))
2710 rq->idle_tokens = -ret;
2711 deactivate_task(next, rq);
2712 list_add_tail(&next->run_list, &rq->hold_queue);
2713 next->state |= TASK_ONHOLD;
2719 if (!rt_task(next) && next->activated > 0) {
2720 unsigned long long delta = now - next->timestamp;
2722 if (next->activated == 1)
2723 delta = delta * (ON_RUNQUEUE_WEIGHT * 128 / 100) / 128;
2725 array = next->array;
2726 dequeue_task(next, array);
2727 recalc_task_prio(next, next->timestamp + delta);
2728 enqueue_task(next, array);
2730 next->activated = 0;
2733 if (test_and_clear_tsk_thread_flag(prev,TIF_NEED_RESCHED))
2735 RCU_qsctr(task_cpu(prev))++;
2737 #ifdef CONFIG_CKRM_CPU_SCHEDULE
2738 if (prev != rq->idle) {
2739 unsigned long long run = now - prev->timestamp;
2740 cpu_demand_event(get_task_local_stat(prev),CPU_DEMAND_DESCHEDULE,run);
2741 update_local_cvt(prev, run);
2745 prev->sleep_avg -= run_time;
2746 if ((long)prev->sleep_avg <= 0) {
2747 prev->sleep_avg = 0;
2748 if (!(HIGH_CREDIT(prev) || LOW_CREDIT(prev)))
2749 prev->interactive_credit--;
2751 add_delay_ts(prev,runcpu_total,prev->timestamp,now);
2752 prev->timestamp = now;
2754 if (likely(prev != next)) {
2755 add_delay_ts(next,waitcpu_total,next->timestamp,now);
2756 inc_delay(next,runs);
2757 next->timestamp = now;
2762 prepare_arch_switch(rq, next);
2763 prev = context_switch(rq, prev, next);
2766 finish_task_switch(prev);
2768 spin_unlock_irq(&rq->lock);
2770 reacquire_kernel_lock(current);
2771 preempt_enable_no_resched();
2772 if (test_thread_flag(TIF_NEED_RESCHED))
2776 EXPORT_SYMBOL(schedule);
2778 #ifdef CONFIG_PREEMPT
2780 * this is is the entry point to schedule() from in-kernel preemption
2781 * off of preempt_enable. Kernel preemptions off return from interrupt
2782 * occur there and call schedule directly.
2784 asmlinkage void __sched preempt_schedule(void)
2786 struct thread_info *ti = current_thread_info();
2789 * If there is a non-zero preempt_count or interrupts are disabled,
2790 * we do not want to preempt the current task. Just return..
2792 if (unlikely(ti->preempt_count || irqs_disabled()))
2796 ti->preempt_count = PREEMPT_ACTIVE;
2798 ti->preempt_count = 0;
2800 /* we could miss a preemption opportunity between schedule and now */
2802 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
2806 EXPORT_SYMBOL(preempt_schedule);
2807 #endif /* CONFIG_PREEMPT */
2809 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync, void *key)
2811 task_t *p = curr->task;
2812 return try_to_wake_up(p, mode, sync);
2815 EXPORT_SYMBOL(default_wake_function);
2818 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2819 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2820 * number) then we wake all the non-exclusive tasks and one exclusive task.
2822 * There are circumstances in which we can try to wake a task which has already
2823 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2824 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2826 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2827 int nr_exclusive, int sync, void *key)
2829 struct list_head *tmp, *next;
2831 list_for_each_safe(tmp, next, &q->task_list) {
2834 curr = list_entry(tmp, wait_queue_t, task_list);
2835 flags = curr->flags;
2836 if (curr->func(curr, mode, sync, key) &&
2837 (flags & WQ_FLAG_EXCLUSIVE) &&
2844 * __wake_up - wake up threads blocked on a waitqueue.
2846 * @mode: which threads
2847 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2849 void fastcall __wake_up(wait_queue_head_t *q, unsigned int mode,
2850 int nr_exclusive, void *key)
2852 unsigned long flags;
2854 spin_lock_irqsave(&q->lock, flags);
2855 __wake_up_common(q, mode, nr_exclusive, 0, key);
2856 spin_unlock_irqrestore(&q->lock, flags);
2859 EXPORT_SYMBOL(__wake_up);
2862 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2864 void fastcall __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
2866 __wake_up_common(q, mode, 1, 0, NULL);
2870 * __wake_up - sync- wake up threads blocked on a waitqueue.
2872 * @mode: which threads
2873 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2875 * The sync wakeup differs that the waker knows that it will schedule
2876 * away soon, so while the target thread will be woken up, it will not
2877 * be migrated to another CPU - ie. the two threads are 'synchronized'
2878 * with each other. This can prevent needless bouncing between CPUs.
2880 * On UP it can prevent extra preemption.
2882 void fastcall __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2884 unsigned long flags;
2890 if (unlikely(!nr_exclusive))
2893 spin_lock_irqsave(&q->lock, flags);
2894 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
2895 spin_unlock_irqrestore(&q->lock, flags);
2897 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2899 void fastcall complete(struct completion *x)
2901 unsigned long flags;
2903 spin_lock_irqsave(&x->wait.lock, flags);
2905 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2907 spin_unlock_irqrestore(&x->wait.lock, flags);
2909 EXPORT_SYMBOL(complete);
2911 void fastcall complete_all(struct completion *x)
2913 unsigned long flags;
2915 spin_lock_irqsave(&x->wait.lock, flags);
2916 x->done += UINT_MAX/2;
2917 __wake_up_common(&x->wait, TASK_UNINTERRUPTIBLE | TASK_INTERRUPTIBLE,
2919 spin_unlock_irqrestore(&x->wait.lock, flags);
2921 EXPORT_SYMBOL(complete_all);
2923 void fastcall __sched wait_for_completion(struct completion *x)
2926 spin_lock_irq(&x->wait.lock);
2928 DECLARE_WAITQUEUE(wait, current);
2930 wait.flags |= WQ_FLAG_EXCLUSIVE;
2931 __add_wait_queue_tail(&x->wait, &wait);
2933 __set_current_state(TASK_UNINTERRUPTIBLE);
2934 spin_unlock_irq(&x->wait.lock);
2936 spin_lock_irq(&x->wait.lock);
2938 __remove_wait_queue(&x->wait, &wait);
2941 spin_unlock_irq(&x->wait.lock);
2943 EXPORT_SYMBOL(wait_for_completion);
2945 #define SLEEP_ON_VAR \
2946 unsigned long flags; \
2947 wait_queue_t wait; \
2948 init_waitqueue_entry(&wait, current);
2950 #define SLEEP_ON_HEAD \
2951 spin_lock_irqsave(&q->lock,flags); \
2952 __add_wait_queue(q, &wait); \
2953 spin_unlock(&q->lock);
2955 #define SLEEP_ON_TAIL \
2956 spin_lock_irq(&q->lock); \
2957 __remove_wait_queue(q, &wait); \
2958 spin_unlock_irqrestore(&q->lock, flags);
2960 #define SLEEP_ON_BKLCHECK \
2961 if (unlikely(!kernel_locked()) && \
2962 sleep_on_bkl_warnings < 10) { \
2963 sleep_on_bkl_warnings++; \
2967 static int sleep_on_bkl_warnings;
2969 void fastcall __sched interruptible_sleep_on(wait_queue_head_t *q)
2975 current->state = TASK_INTERRUPTIBLE;
2982 EXPORT_SYMBOL(interruptible_sleep_on);
2984 long fastcall __sched interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2990 current->state = TASK_INTERRUPTIBLE;
2993 timeout = schedule_timeout(timeout);
2999 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3001 long fastcall __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3007 current->state = TASK_UNINTERRUPTIBLE;
3010 timeout = schedule_timeout(timeout);
3016 EXPORT_SYMBOL(sleep_on_timeout);
3018 void set_user_nice(task_t *p, long nice)
3020 unsigned long flags;
3021 prio_array_t *array;
3023 int old_prio, new_prio, delta;
3025 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3028 * We have to be careful, if called from sys_setpriority(),
3029 * the task might be in the middle of scheduling on another CPU.
3031 rq = task_rq_lock(p, &flags);
3033 * The RT priorities are set via setscheduler(), but we still
3034 * allow the 'normal' nice value to be set - but as expected
3035 * it wont have any effect on scheduling until the task is
3039 p->static_prio = NICE_TO_PRIO(nice);
3044 dequeue_task(p, array);
3047 new_prio = NICE_TO_PRIO(nice);
3048 delta = new_prio - old_prio;
3049 p->static_prio = NICE_TO_PRIO(nice);
3053 enqueue_task(p, array);
3055 * If the task increased its priority or is running and
3056 * lowered its priority, then reschedule its CPU:
3058 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3059 resched_task(rq->curr);
3062 task_rq_unlock(rq, &flags);
3065 EXPORT_SYMBOL(set_user_nice);
3067 #ifdef __ARCH_WANT_SYS_NICE
3070 * sys_nice - change the priority of the current process.
3071 * @increment: priority increment
3073 * sys_setpriority is a more generic, but much slower function that
3074 * does similar things.
3076 asmlinkage long sys_nice(int increment)
3082 * Setpriority might change our priority at the same moment.
3083 * We don't have to worry. Conceptually one call occurs first
3084 * and we have a single winner.
3086 if (increment < 0) {
3087 if (!capable(CAP_SYS_NICE))
3089 if (increment < -40)
3095 nice = PRIO_TO_NICE(current->static_prio) + increment;
3101 retval = security_task_setnice(current, nice);
3105 set_user_nice(current, nice);
3112 * task_prio - return the priority value of a given task.
3113 * @p: the task in question.
3115 * This is the priority value as seen by users in /proc.
3116 * RT tasks are offset by -200. Normal tasks are centered
3117 * around 0, value goes from -16 to +15.
3119 int task_prio(const task_t *p)
3121 return p->prio - MAX_RT_PRIO;
3125 * task_nice - return the nice value of a given task.
3126 * @p: the task in question.
3128 int task_nice(const task_t *p)
3130 return TASK_NICE(p);
3133 EXPORT_SYMBOL(task_nice);
3136 * idle_cpu - is a given cpu idle currently?
3137 * @cpu: the processor in question.
3139 int idle_cpu(int cpu)
3141 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
3144 EXPORT_SYMBOL_GPL(idle_cpu);
3147 * find_process_by_pid - find a process with a matching PID value.
3148 * @pid: the pid in question.
3150 static inline task_t *find_process_by_pid(pid_t pid)
3152 return pid ? find_task_by_pid(pid) : current;
3155 /* Actually do priority change: must hold rq lock. */
3156 static void __setscheduler(struct task_struct *p, int policy, int prio)
3160 p->rt_priority = prio;
3161 if (policy != SCHED_NORMAL)
3162 p->prio = MAX_USER_RT_PRIO-1 - p->rt_priority;
3164 p->prio = p->static_prio;
3168 * setscheduler - change the scheduling policy and/or RT priority of a thread.
3170 static int setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3172 struct sched_param lp;
3173 int retval = -EINVAL;
3175 prio_array_t *array;
3176 unsigned long flags;
3180 if (!param || pid < 0)
3184 if (copy_from_user(&lp, param, sizeof(struct sched_param)))
3188 * We play safe to avoid deadlocks.
3190 read_lock_irq(&tasklist_lock);
3192 p = find_process_by_pid(pid);
3196 goto out_unlock_tasklist;
3199 * To be able to change p->policy safely, the apropriate
3200 * runqueue lock must be held.
3202 rq = task_rq_lock(p, &flags);
3208 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3209 policy != SCHED_NORMAL)
3214 * Valid priorities for SCHED_FIFO and SCHED_RR are
3215 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL is 0.
3218 if (lp.sched_priority < 0 || lp.sched_priority > MAX_USER_RT_PRIO-1)
3220 if ((policy == SCHED_NORMAL) != (lp.sched_priority == 0))
3224 if ((policy == SCHED_FIFO || policy == SCHED_RR) &&
3225 !capable(CAP_SYS_NICE))
3227 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3228 !capable(CAP_SYS_NICE))
3231 retval = security_task_setscheduler(p, policy, &lp);
3237 deactivate_task(p, task_rq(p));
3240 __setscheduler(p, policy, lp.sched_priority);
3242 __activate_task(p, task_rq(p));
3244 * Reschedule if we are currently running on this runqueue and
3245 * our priority decreased, or if we are not currently running on
3246 * this runqueue and our priority is higher than the current's
3248 if (task_running(rq, p)) {
3249 if (p->prio > oldprio)
3250 resched_task(rq->curr);
3251 } else if (TASK_PREEMPTS_CURR(p, rq))
3252 resched_task(rq->curr);
3256 task_rq_unlock(rq, &flags);
3257 out_unlock_tasklist:
3258 read_unlock_irq(&tasklist_lock);
3265 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3266 * @pid: the pid in question.
3267 * @policy: new policy
3268 * @param: structure containing the new RT priority.
3270 asmlinkage long sys_sched_setscheduler(pid_t pid, int policy,
3271 struct sched_param __user *param)
3273 return setscheduler(pid, policy, param);
3277 * sys_sched_setparam - set/change the RT priority of a thread
3278 * @pid: the pid in question.
3279 * @param: structure containing the new RT priority.
3281 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
3283 return setscheduler(pid, -1, param);
3287 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3288 * @pid: the pid in question.
3290 asmlinkage long sys_sched_getscheduler(pid_t pid)
3292 int retval = -EINVAL;
3299 read_lock(&tasklist_lock);
3300 p = find_process_by_pid(pid);
3302 retval = security_task_getscheduler(p);
3306 read_unlock(&tasklist_lock);
3313 * sys_sched_getscheduler - get the RT priority of a thread
3314 * @pid: the pid in question.
3315 * @param: structure containing the RT priority.
3317 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
3319 struct sched_param lp;
3320 int retval = -EINVAL;
3323 if (!param || pid < 0)
3326 read_lock(&tasklist_lock);
3327 p = find_process_by_pid(pid);
3332 retval = security_task_getscheduler(p);
3336 lp.sched_priority = p->rt_priority;
3337 read_unlock(&tasklist_lock);
3340 * This one might sleep, we cannot do it with a spinlock held ...
3342 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3348 read_unlock(&tasklist_lock);
3353 * sys_sched_setaffinity - set the cpu affinity of a process
3354 * @pid: pid of the process
3355 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3356 * @user_mask_ptr: user-space pointer to the new cpu mask
3358 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
3359 unsigned long __user *user_mask_ptr)
3365 if (len < sizeof(new_mask))
3368 if (copy_from_user(&new_mask, user_mask_ptr, sizeof(new_mask)))
3372 read_lock(&tasklist_lock);
3374 p = find_process_by_pid(pid);
3376 read_unlock(&tasklist_lock);
3377 unlock_cpu_hotplug();
3382 * It is not safe to call set_cpus_allowed with the
3383 * tasklist_lock held. We will bump the task_struct's
3384 * usage count and then drop tasklist_lock.
3387 read_unlock(&tasklist_lock);
3390 if ((current->euid != p->euid) && (current->euid != p->uid) &&
3391 !capable(CAP_SYS_NICE))
3394 retval = set_cpus_allowed(p, new_mask);
3398 unlock_cpu_hotplug();
3403 * Represents all cpu's present in the system
3404 * In systems capable of hotplug, this map could dynamically grow
3405 * as new cpu's are detected in the system via any platform specific
3406 * method, such as ACPI for e.g.
3409 cpumask_t cpu_present_map;
3410 EXPORT_SYMBOL(cpu_present_map);
3413 cpumask_t cpu_online_map = CPU_MASK_ALL;
3414 cpumask_t cpu_possible_map = CPU_MASK_ALL;
3418 * sys_sched_getaffinity - get the cpu affinity of a process
3419 * @pid: pid of the process
3420 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3421 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3423 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
3424 unsigned long __user *user_mask_ptr)
3426 unsigned int real_len;
3431 real_len = sizeof(mask);
3436 read_lock(&tasklist_lock);
3439 p = find_process_by_pid(pid);
3444 cpus_and(mask, p->cpus_allowed, cpu_possible_map);
3447 read_unlock(&tasklist_lock);
3448 unlock_cpu_hotplug();
3451 if (copy_to_user(user_mask_ptr, &mask, real_len))
3457 * sys_sched_yield - yield the current processor to other threads.
3459 * this function yields the current CPU by moving the calling thread
3460 * to the expired array. If there are no other threads running on this
3461 * CPU then this function will return.
3463 asmlinkage long sys_sched_yield(void)
3465 runqueue_t *rq = this_rq_lock();
3466 prio_array_t *array = current->array;
3467 prio_array_t *target = rq_expired(current,rq);
3470 * We implement yielding by moving the task into the expired
3473 * (special rule: RT tasks will just roundrobin in the active
3476 if (unlikely(rt_task(current)))
3477 target = rq_active(current,rq);
3479 dequeue_task(current, array);
3480 enqueue_task(current, target);
3483 * Since we are going to call schedule() anyway, there's
3484 * no need to preempt or enable interrupts:
3486 _raw_spin_unlock(&rq->lock);
3487 preempt_enable_no_resched();
3494 void __sched __cond_resched(void)
3496 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
3497 __might_sleep(__FILE__, __LINE__, 0);
3500 * The system_state check is somewhat ugly but we might be
3501 * called during early boot when we are not yet ready to reschedule.
3503 if (need_resched() && system_state >= SYSTEM_BOOTING_SCHEDULER_OK) {
3504 set_current_state(TASK_RUNNING);
3509 EXPORT_SYMBOL(__cond_resched);
3511 void __sched __cond_resched_lock(spinlock_t * lock)
3513 if (need_resched()) {
3514 _raw_spin_unlock(lock);
3515 preempt_enable_no_resched();
3516 set_current_state(TASK_RUNNING);
3522 EXPORT_SYMBOL(__cond_resched_lock);
3525 * yield - yield the current processor to other threads.
3527 * this is a shortcut for kernel-space yielding - it marks the
3528 * thread runnable and calls sys_sched_yield().
3530 void __sched yield(void)
3532 set_current_state(TASK_RUNNING);
3536 EXPORT_SYMBOL(yield);
3539 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3540 * that process accounting knows that this is a task in IO wait state.
3542 * But don't do that if it is a deliberate, throttling IO wait (this task
3543 * has set its backing_dev_info: the queue against which it should throttle)
3545 void __sched io_schedule(void)
3547 struct runqueue *rq = this_rq();
3548 def_delay_var(dstart);
3550 start_delay_set(dstart,PF_IOWAIT);
3551 atomic_inc(&rq->nr_iowait);
3553 atomic_dec(&rq->nr_iowait);
3554 add_io_delay(dstart);
3557 EXPORT_SYMBOL(io_schedule);
3559 long __sched io_schedule_timeout(long timeout)
3561 struct runqueue *rq = this_rq();
3563 def_delay_var(dstart);
3565 start_delay_set(dstart,PF_IOWAIT);
3566 atomic_inc(&rq->nr_iowait);
3567 ret = schedule_timeout(timeout);
3568 atomic_dec(&rq->nr_iowait);
3569 add_io_delay(dstart);
3574 * sys_sched_get_priority_max - return maximum RT priority.
3575 * @policy: scheduling class.
3577 * this syscall returns the maximum rt_priority that can be used
3578 * by a given scheduling class.
3580 asmlinkage long sys_sched_get_priority_max(int policy)
3587 ret = MAX_USER_RT_PRIO-1;
3597 * sys_sched_get_priority_min - return minimum RT priority.
3598 * @policy: scheduling class.
3600 * this syscall returns the minimum rt_priority that can be used
3601 * by a given scheduling class.
3603 asmlinkage long sys_sched_get_priority_min(int policy)
3619 * sys_sched_rr_get_interval - return the default timeslice of a process.
3620 * @pid: pid of the process.
3621 * @interval: userspace pointer to the timeslice value.
3623 * this syscall writes the default timeslice value of a given process
3624 * into the user-space timespec buffer. A value of '0' means infinity.
3627 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
3629 int retval = -EINVAL;
3637 read_lock(&tasklist_lock);
3638 p = find_process_by_pid(pid);
3642 retval = security_task_getscheduler(p);
3646 jiffies_to_timespec(p->policy & SCHED_FIFO ?
3647 0 : task_timeslice(p), &t);
3648 read_unlock(&tasklist_lock);
3649 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
3653 read_unlock(&tasklist_lock);
3657 static inline struct task_struct *eldest_child(struct task_struct *p)
3659 if (list_empty(&p->children)) return NULL;
3660 return list_entry(p->children.next,struct task_struct,sibling);
3663 static inline struct task_struct *older_sibling(struct task_struct *p)
3665 if (p->sibling.prev==&p->parent->children) return NULL;
3666 return list_entry(p->sibling.prev,struct task_struct,sibling);
3669 static inline struct task_struct *younger_sibling(struct task_struct *p)
3671 if (p->sibling.next==&p->parent->children) return NULL;
3672 return list_entry(p->sibling.next,struct task_struct,sibling);
3675 static void show_task(task_t * p)
3679 unsigned long free = 0;
3680 static const char *stat_nam[] = { "R", "S", "D", "T", "Z", "W" };
3682 printk("%-13.13s ", p->comm);
3683 state = p->state ? __ffs(p->state) + 1 : 0;
3684 if (state < ARRAY_SIZE(stat_nam))
3685 printk(stat_nam[state]);
3688 #if (BITS_PER_LONG == 32)
3689 if (state == TASK_RUNNING)
3690 printk(" running ");
3692 printk(" %08lX ", thread_saved_pc(p));
3694 if (state == TASK_RUNNING)
3695 printk(" running task ");
3697 printk(" %016lx ", thread_saved_pc(p));
3699 #ifdef CONFIG_DEBUG_STACK_USAGE
3701 unsigned long * n = (unsigned long *) (p->thread_info+1);
3704 free = (unsigned long) n - (unsigned long)(p->thread_info+1);
3707 printk("%5lu %5d %6d ", free, p->pid, p->parent->pid);
3708 if ((relative = eldest_child(p)))
3709 printk("%5d ", relative->pid);
3712 if ((relative = younger_sibling(p)))
3713 printk("%7d", relative->pid);
3716 if ((relative = older_sibling(p)))
3717 printk(" %5d", relative->pid);
3721 printk(" (L-TLB)\n");
3723 printk(" (NOTLB)\n");
3725 if (state != TASK_RUNNING)
3726 show_stack(p, NULL);
3729 void show_state(void)
3733 #if (BITS_PER_LONG == 32)
3736 printk(" task PC pid father child younger older\n");
3740 printk(" task PC pid father child younger older\n");
3742 read_lock(&tasklist_lock);
3743 do_each_thread(g, p) {
3745 * reset the NMI-timeout, listing all files on a slow
3746 * console might take alot of time:
3748 touch_nmi_watchdog();
3750 } while_each_thread(g, p);
3752 read_unlock(&tasklist_lock);
3755 EXPORT_SYMBOL_GPL(show_state);
3757 void __devinit init_idle(task_t *idle, int cpu)
3759 runqueue_t *idle_rq = cpu_rq(cpu), *rq = cpu_rq(task_cpu(idle));
3760 unsigned long flags;
3762 local_irq_save(flags);
3763 double_rq_lock(idle_rq, rq);
3765 idle_rq->curr = idle_rq->idle = idle;
3766 deactivate_task(idle, rq);
3768 idle->prio = MAX_PRIO;
3769 idle->state = TASK_RUNNING;
3770 set_task_cpu(idle, cpu);
3771 double_rq_unlock(idle_rq, rq);
3772 set_tsk_need_resched(idle);
3773 local_irq_restore(flags);
3775 /* Set the preempt count _outside_ the spinlocks! */
3776 #ifdef CONFIG_PREEMPT
3777 idle->thread_info->preempt_count = (idle->lock_depth >= 0);
3779 idle->thread_info->preempt_count = 0;
3784 * In a system that switches off the HZ timer nohz_cpu_mask
3785 * indicates which cpus entered this state. This is used
3786 * in the rcu update to wait only for active cpus. For system
3787 * which do not switch off the HZ timer nohz_cpu_mask should
3788 * always be CPU_MASK_NONE.
3790 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
3794 * This is how migration works:
3796 * 1) we queue a migration_req_t structure in the source CPU's
3797 * runqueue and wake up that CPU's migration thread.
3798 * 2) we down() the locked semaphore => thread blocks.
3799 * 3) migration thread wakes up (implicitly it forces the migrated
3800 * thread off the CPU)
3801 * 4) it gets the migration request and checks whether the migrated
3802 * task is still in the wrong runqueue.
3803 * 5) if it's in the wrong runqueue then the migration thread removes
3804 * it and puts it into the right queue.
3805 * 6) migration thread up()s the semaphore.
3806 * 7) we wake up and the migration is done.
3810 * Change a given task's CPU affinity. Migrate the thread to a
3811 * proper CPU and schedule it away if the CPU it's executing on
3812 * is removed from the allowed bitmask.
3814 * NOTE: the caller must have a valid reference to the task, the
3815 * task must not exit() & deallocate itself prematurely. The
3816 * call is not atomic; no spinlocks may be held.
3818 int set_cpus_allowed(task_t *p, cpumask_t new_mask)
3820 unsigned long flags;
3822 migration_req_t req;
3825 rq = task_rq_lock(p, &flags);
3826 if (!cpus_intersects(new_mask, cpu_online_map)) {
3831 p->cpus_allowed = new_mask;
3832 /* Can the task run on the task's current CPU? If so, we're done */
3833 if (cpu_isset(task_cpu(p), new_mask))
3836 if (migrate_task(p, any_online_cpu(new_mask), &req)) {
3837 /* Need help from migration thread: drop lock and wait. */
3838 task_rq_unlock(rq, &flags);
3839 wake_up_process(rq->migration_thread);
3840 wait_for_completion(&req.done);
3841 tlb_migrate_finish(p->mm);
3845 task_rq_unlock(rq, &flags);
3849 EXPORT_SYMBOL_GPL(set_cpus_allowed);
3852 * Move (not current) task off this cpu, onto dest cpu. We're doing
3853 * this because either it can't run here any more (set_cpus_allowed()
3854 * away from this CPU, or CPU going down), or because we're
3855 * attempting to rebalance this task on exec (sched_balance_exec).
3857 * So we race with normal scheduler movements, but that's OK, as long
3858 * as the task is no longer on this CPU.
3860 static void __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
3862 runqueue_t *rq_dest, *rq_src;
3864 if (unlikely(cpu_is_offline(dest_cpu)))
3867 rq_src = cpu_rq(src_cpu);
3868 rq_dest = cpu_rq(dest_cpu);
3870 double_rq_lock(rq_src, rq_dest);
3871 /* Already moved. */
3872 if (task_cpu(p) != src_cpu)
3874 /* Affinity changed (again). */
3875 if (!cpu_isset(dest_cpu, p->cpus_allowed))
3878 set_task_cpu(p, dest_cpu);
3881 * Sync timestamp with rq_dest's before activating.
3882 * The same thing could be achieved by doing this step
3883 * afterwards, and pretending it was a local activate.
3884 * This way is cleaner and logically correct.
3886 p->timestamp = p->timestamp - rq_src->timestamp_last_tick
3887 + rq_dest->timestamp_last_tick;
3888 deactivate_task(p, rq_src);
3889 activate_task(p, rq_dest, 0);
3890 if (TASK_PREEMPTS_CURR(p, rq_dest))
3891 resched_task(rq_dest->curr);
3895 double_rq_unlock(rq_src, rq_dest);
3899 * migration_thread - this is a highprio system thread that performs
3900 * thread migration by bumping thread off CPU then 'pushing' onto
3903 static int migration_thread(void * data)
3906 int cpu = (long)data;
3909 BUG_ON(rq->migration_thread != current);
3911 set_current_state(TASK_INTERRUPTIBLE);
3912 while (!kthread_should_stop()) {
3913 struct list_head *head;
3914 migration_req_t *req;
3916 if (current->flags & PF_FREEZE)
3917 refrigerator(PF_FREEZE);
3919 spin_lock_irq(&rq->lock);
3921 if (cpu_is_offline(cpu)) {
3922 spin_unlock_irq(&rq->lock);
3926 if (rq->active_balance) {
3927 #ifndef CONFIG_CKRM_CPU_SCHEDULE
3928 active_load_balance(rq, cpu);
3930 rq->active_balance = 0;
3933 head = &rq->migration_queue;
3935 if (list_empty(head)) {
3936 spin_unlock_irq(&rq->lock);
3938 set_current_state(TASK_INTERRUPTIBLE);
3941 req = list_entry(head->next, migration_req_t, list);
3942 list_del_init(head->next);
3944 if (req->type == REQ_MOVE_TASK) {
3945 spin_unlock(&rq->lock);
3946 __migrate_task(req->task, smp_processor_id(),
3949 } else if (req->type == REQ_SET_DOMAIN) {
3951 spin_unlock_irq(&rq->lock);
3953 spin_unlock_irq(&rq->lock);
3957 complete(&req->done);
3959 __set_current_state(TASK_RUNNING);
3963 /* Wait for kthread_stop */
3964 set_current_state(TASK_INTERRUPTIBLE);
3965 while (!kthread_should_stop()) {
3967 set_current_state(TASK_INTERRUPTIBLE);
3969 __set_current_state(TASK_RUNNING);
3973 #ifdef CONFIG_HOTPLUG_CPU
3974 /* migrate_all_tasks - function to migrate all tasks from the dead cpu. */
3975 static void migrate_all_tasks(int src_cpu)
3977 struct task_struct *tsk, *t;
3981 write_lock_irq(&tasklist_lock);
3983 /* watch out for per node tasks, let's stay on this node */
3984 node = cpu_to_node(src_cpu);
3986 do_each_thread(t, tsk) {
3991 if (task_cpu(tsk) != src_cpu)
3994 /* Figure out where this task should go (attempting to
3995 * keep it on-node), and check if it can be migrated
3996 * as-is. NOTE that kernel threads bound to more than
3997 * one online cpu will be migrated. */
3998 mask = node_to_cpumask(node);
3999 cpus_and(mask, mask, tsk->cpus_allowed);
4000 dest_cpu = any_online_cpu(mask);
4001 if (dest_cpu == NR_CPUS)
4002 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4003 if (dest_cpu == NR_CPUS) {
4004 cpus_setall(tsk->cpus_allowed);
4005 dest_cpu = any_online_cpu(tsk->cpus_allowed);
4007 /* Don't tell them about moving exiting tasks
4008 or kernel threads (both mm NULL), since
4009 they never leave kernel. */
4010 if (tsk->mm && printk_ratelimit())
4011 printk(KERN_INFO "process %d (%s) no "
4012 "longer affine to cpu%d\n",
4013 tsk->pid, tsk->comm, src_cpu);
4016 __migrate_task(tsk, src_cpu, dest_cpu);
4017 } while_each_thread(t, tsk);
4019 write_unlock_irq(&tasklist_lock);
4022 /* Schedules idle task to be the next runnable task on current CPU.
4023 * It does so by boosting its priority to highest possible and adding it to
4024 * the _front_ of runqueue. Used by CPU offline code.
4026 void sched_idle_next(void)
4028 int cpu = smp_processor_id();
4029 runqueue_t *rq = this_rq();
4030 struct task_struct *p = rq->idle;
4031 unsigned long flags;
4033 /* cpu has to be offline */
4034 BUG_ON(cpu_online(cpu));
4036 /* Strictly not necessary since rest of the CPUs are stopped by now
4037 * and interrupts disabled on current cpu.
4039 spin_lock_irqsave(&rq->lock, flags);
4041 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4042 /* Add idle task to _front_ of it's priority queue */
4043 __activate_idle_task(p, rq);
4045 spin_unlock_irqrestore(&rq->lock, flags);
4047 #endif /* CONFIG_HOTPLUG_CPU */
4050 * migration_call - callback that gets triggered when a CPU is added.
4051 * Here we can start up the necessary migration thread for the new CPU.
4053 static int migration_call(struct notifier_block *nfb, unsigned long action,
4056 int cpu = (long)hcpu;
4057 struct task_struct *p;
4058 struct runqueue *rq;
4059 unsigned long flags;
4062 case CPU_UP_PREPARE:
4063 p = kthread_create(migration_thread, hcpu, "migration/%d",cpu);
4066 p->flags |= PF_NOFREEZE;
4067 kthread_bind(p, cpu);
4068 /* Must be high prio: stop_machine expects to yield to it. */
4069 rq = task_rq_lock(p, &flags);
4070 __setscheduler(p, SCHED_FIFO, MAX_RT_PRIO-1);
4071 task_rq_unlock(rq, &flags);
4072 cpu_rq(cpu)->migration_thread = p;
4075 /* Strictly unneccessary, as first user will wake it. */
4076 wake_up_process(cpu_rq(cpu)->migration_thread);
4078 #ifdef CONFIG_HOTPLUG_CPU
4079 case CPU_UP_CANCELED:
4080 /* Unbind it from offline cpu so it can run. Fall thru. */
4081 kthread_bind(cpu_rq(cpu)->migration_thread,smp_processor_id());
4082 kthread_stop(cpu_rq(cpu)->migration_thread);
4083 cpu_rq(cpu)->migration_thread = NULL;
4086 migrate_all_tasks(cpu);
4088 kthread_stop(rq->migration_thread);
4089 rq->migration_thread = NULL;
4090 /* Idle task back to normal (off runqueue, low prio) */
4091 rq = task_rq_lock(rq->idle, &flags);
4092 deactivate_task(rq->idle, rq);
4093 rq->idle->static_prio = MAX_PRIO;
4094 __setscheduler(rq->idle, SCHED_NORMAL, 0);
4095 task_rq_unlock(rq, &flags);
4096 BUG_ON(rq->nr_running != 0);
4098 /* No need to migrate the tasks: it was best-effort if
4099 * they didn't do lock_cpu_hotplug(). Just wake up
4100 * the requestors. */
4101 spin_lock_irq(&rq->lock);
4102 while (!list_empty(&rq->migration_queue)) {
4103 migration_req_t *req;
4104 req = list_entry(rq->migration_queue.next,
4105 migration_req_t, list);
4106 BUG_ON(req->type != REQ_MOVE_TASK);
4107 list_del_init(&req->list);
4108 complete(&req->done);
4110 spin_unlock_irq(&rq->lock);
4117 /* Register at highest priority so that task migration (migrate_all_tasks)
4118 * happens before everything else.
4120 static struct notifier_block __devinitdata migration_notifier = {
4121 .notifier_call = migration_call,
4125 int __init migration_init(void)
4127 void *cpu = (void *)(long)smp_processor_id();
4128 /* Start one for boot CPU. */
4129 migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4130 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4131 register_cpu_notifier(&migration_notifier);
4137 * The 'big kernel lock'
4139 * This spinlock is taken and released recursively by lock_kernel()
4140 * and unlock_kernel(). It is transparently dropped and reaquired
4141 * over schedule(). It is used to protect legacy code that hasn't
4142 * been migrated to a proper locking design yet.
4144 * Don't use in new code.
4146 * Note: spinlock debugging needs this even on !CONFIG_SMP.
4148 spinlock_t kernel_flag __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
4149 EXPORT_SYMBOL(kernel_flag);
4152 /* Attach the domain 'sd' to 'cpu' as its base domain */
4153 void cpu_attach_domain(struct sched_domain *sd, int cpu)
4155 migration_req_t req;
4156 unsigned long flags;
4157 runqueue_t *rq = cpu_rq(cpu);
4162 spin_lock_irqsave(&rq->lock, flags);
4164 if (cpu == smp_processor_id() || !cpu_online(cpu)) {
4167 init_completion(&req.done);
4168 req.type = REQ_SET_DOMAIN;
4170 list_add(&req.list, &rq->migration_queue);
4174 spin_unlock_irqrestore(&rq->lock, flags);
4177 wake_up_process(rq->migration_thread);
4178 wait_for_completion(&req.done);
4181 unlock_cpu_hotplug();
4184 #ifdef ARCH_HAS_SCHED_DOMAIN
4185 extern void __init arch_init_sched_domains(void);
4187 static struct sched_group sched_group_cpus[NR_CPUS];
4188 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
4190 static struct sched_group sched_group_nodes[MAX_NUMNODES];
4191 static DEFINE_PER_CPU(struct sched_domain, node_domains);
4192 static void __init arch_init_sched_domains(void)
4195 struct sched_group *first_node = NULL, *last_node = NULL;
4197 /* Set up domains */
4199 int node = cpu_to_node(i);
4200 cpumask_t nodemask = node_to_cpumask(node);
4201 struct sched_domain *node_sd = &per_cpu(node_domains, i);
4202 struct sched_domain *cpu_sd = &per_cpu(cpu_domains, i);
4204 *node_sd = SD_NODE_INIT;
4205 node_sd->span = cpu_possible_map;
4206 node_sd->groups = &sched_group_nodes[cpu_to_node(i)];
4208 *cpu_sd = SD_CPU_INIT;
4209 cpus_and(cpu_sd->span, nodemask, cpu_possible_map);
4210 cpu_sd->groups = &sched_group_cpus[i];
4211 cpu_sd->parent = node_sd;
4215 for (i = 0; i < MAX_NUMNODES; i++) {
4216 cpumask_t tmp = node_to_cpumask(i);
4218 struct sched_group *first_cpu = NULL, *last_cpu = NULL;
4219 struct sched_group *node = &sched_group_nodes[i];
4222 cpus_and(nodemask, tmp, cpu_possible_map);
4224 if (cpus_empty(nodemask))
4227 node->cpumask = nodemask;
4228 node->cpu_power = SCHED_LOAD_SCALE * cpus_weight(node->cpumask);
4230 for_each_cpu_mask(j, node->cpumask) {
4231 struct sched_group *cpu = &sched_group_cpus[j];
4233 cpus_clear(cpu->cpumask);
4234 cpu_set(j, cpu->cpumask);
4235 cpu->cpu_power = SCHED_LOAD_SCALE;
4240 last_cpu->next = cpu;
4243 last_cpu->next = first_cpu;
4248 last_node->next = node;
4251 last_node->next = first_node;
4255 struct sched_domain *cpu_sd = &per_cpu(cpu_domains, i);
4256 cpu_attach_domain(cpu_sd, i);
4260 #else /* !CONFIG_NUMA */
4261 static void __init arch_init_sched_domains(void)
4264 struct sched_group *first_cpu = NULL, *last_cpu = NULL;
4266 /* Set up domains */
4268 struct sched_domain *cpu_sd = &per_cpu(cpu_domains, i);
4270 *cpu_sd = SD_CPU_INIT;
4271 cpu_sd->span = cpu_possible_map;
4272 cpu_sd->groups = &sched_group_cpus[i];
4275 /* Set up CPU groups */
4276 for_each_cpu_mask(i, cpu_possible_map) {
4277 struct sched_group *cpu = &sched_group_cpus[i];
4279 cpus_clear(cpu->cpumask);
4280 cpu_set(i, cpu->cpumask);
4281 cpu->cpu_power = SCHED_LOAD_SCALE;
4286 last_cpu->next = cpu;
4289 last_cpu->next = first_cpu;
4291 mb(); /* domains were modified outside the lock */
4293 struct sched_domain *cpu_sd = &per_cpu(cpu_domains, i);
4294 cpu_attach_domain(cpu_sd, i);
4298 #endif /* CONFIG_NUMA */
4299 #endif /* ARCH_HAS_SCHED_DOMAIN */
4301 #define SCHED_DOMAIN_DEBUG
4302 #ifdef SCHED_DOMAIN_DEBUG
4303 void sched_domain_debug(void)
4308 runqueue_t *rq = cpu_rq(i);
4309 struct sched_domain *sd;
4314 printk(KERN_DEBUG "CPU%d: %s\n",
4315 i, (cpu_online(i) ? " online" : "offline"));
4320 struct sched_group *group = sd->groups;
4321 cpumask_t groupmask;
4323 cpumask_scnprintf(str, NR_CPUS, sd->span);
4324 cpus_clear(groupmask);
4327 for (j = 0; j < level + 1; j++)
4329 printk("domain %d: span %s\n", level, str);
4331 if (!cpu_isset(i, sd->span))
4332 printk(KERN_DEBUG "ERROR domain->span does not contain CPU%d\n", i);
4333 if (!cpu_isset(i, group->cpumask))
4334 printk(KERN_DEBUG "ERROR domain->groups does not contain CPU%d\n", i);
4335 if (!group->cpu_power)
4336 printk(KERN_DEBUG "ERROR domain->cpu_power not set\n");
4339 for (j = 0; j < level + 2; j++)
4344 printk(" ERROR: NULL");
4348 if (!cpus_weight(group->cpumask))
4349 printk(" ERROR empty group:");
4351 if (cpus_intersects(groupmask, group->cpumask))
4352 printk(" ERROR repeated CPUs:");
4354 cpus_or(groupmask, groupmask, group->cpumask);
4356 cpumask_scnprintf(str, NR_CPUS, group->cpumask);
4359 group = group->next;
4360 } while (group != sd->groups);
4363 if (!cpus_equal(sd->span, groupmask))
4364 printk(KERN_DEBUG "ERROR groups don't span domain->span\n");
4370 if (!cpus_subset(groupmask, sd->span))
4371 printk(KERN_DEBUG "ERROR parent span is not a superset of domain->span\n");
4378 #define sched_domain_debug() {}
4381 void __init sched_init_smp(void)
4383 arch_init_sched_domains();
4384 sched_domain_debug();
4387 void __init sched_init_smp(void)
4390 #endif /* CONFIG_SMP */
4392 int in_sched_functions(unsigned long addr)
4394 /* Linker adds these: start and end of __sched functions */
4395 extern char __sched_text_start[], __sched_text_end[];
4396 return addr >= (unsigned long)__sched_text_start
4397 && addr < (unsigned long)__sched_text_end;
4400 void __init sched_init(void)
4404 #ifndef CONFIG_CKRM_CPU_SCHEDULE
4409 /* Set up an initial dummy domain for early boot */
4410 static struct sched_domain sched_domain_init;
4411 static struct sched_group sched_group_init;
4413 memset(&sched_domain_init, 0, sizeof(struct sched_domain));
4414 sched_domain_init.span = CPU_MASK_ALL;
4415 sched_domain_init.groups = &sched_group_init;
4416 sched_domain_init.last_balance = jiffies;
4417 sched_domain_init.balance_interval = INT_MAX; /* Don't balance */
4419 memset(&sched_group_init, 0, sizeof(struct sched_group));
4420 sched_group_init.cpumask = CPU_MASK_ALL;
4421 sched_group_init.next = &sched_group_init;
4422 sched_group_init.cpu_power = SCHED_LOAD_SCALE;
4427 for (i = 0; i < NR_CPUS; i++) {
4428 #ifndef CONFIG_CKRM_CPU_SCHEDULE
4429 prio_array_t *array;
4432 spin_lock_init(&rq->lock);
4434 #ifndef CONFIG_CKRM_CPU_SCHEDULE
4435 rq->active = rq->arrays;
4436 rq->expired = rq->arrays + 1;
4438 rq->ckrm_cpu_load = 0;
4440 rq->best_expired_prio = MAX_PRIO;
4443 rq->sd = &sched_domain_init;
4445 rq->active_balance = 0;
4447 rq->migration_thread = NULL;
4448 INIT_LIST_HEAD(&rq->migration_queue);
4450 INIT_LIST_HEAD(&rq->hold_queue);
4451 atomic_set(&rq->nr_iowait, 0);
4453 #ifndef CONFIG_CKRM_CPU_SCHEDULE
4454 for (j = 0; j < 2; j++) {
4455 array = rq->arrays + j;
4456 for (k = 0; k < MAX_PRIO; k++) {
4457 INIT_LIST_HEAD(array->queue + k);
4458 __clear_bit(k, array->bitmap);
4460 // delimiter for bitsearch
4461 __set_bit(MAX_PRIO, array->bitmap);
4467 * We have to do a little magic to get the first
4468 * thread right in SMP mode.
4473 set_task_cpu(current, smp_processor_id());
4474 #ifdef CONFIG_CKRM_CPU_SCHEDULE
4475 current->cpu_class = default_cpu_class;
4476 current->array = NULL;
4478 wake_up_forked_process(current);
4481 * The boot idle thread does lazy MMU switching as well:
4483 atomic_inc(&init_mm.mm_count);
4484 enter_lazy_tlb(&init_mm, current);
4487 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4488 void __might_sleep(char *file, int line, int atomic_depth)
4490 #if defined(in_atomic)
4491 static unsigned long prev_jiffy; /* ratelimiting */
4493 #ifndef CONFIG_PREEMPT
4496 if (((in_atomic() != atomic_depth) || irqs_disabled()) &&
4497 system_state == SYSTEM_RUNNING) {
4498 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
4500 prev_jiffy = jiffies;
4501 printk(KERN_ERR "Debug: sleeping function called from invalid"
4502 " context at %s:%d\n", file, line);
4503 printk("in_atomic():%d[expected: %d], irqs_disabled():%d\n",
4504 in_atomic(), atomic_depth, irqs_disabled());
4509 EXPORT_SYMBOL(__might_sleep);
4513 #if defined(CONFIG_SMP) && defined(CONFIG_PREEMPT)
4515 * This could be a long-held lock. If another CPU holds it for a long time,
4516 * and that CPU is not asked to reschedule then *this* CPU will spin on the
4517 * lock for a long time, even if *this* CPU is asked to reschedule.
4519 * So what we do here, in the slow (contended) path is to spin on the lock by
4520 * hand while permitting preemption.
4522 * Called inside preempt_disable().
4524 void __sched __preempt_spin_lock(spinlock_t *lock)
4526 if (preempt_count() > 1) {
4527 _raw_spin_lock(lock);
4532 while (spin_is_locked(lock))
4535 } while (!_raw_spin_trylock(lock));
4538 EXPORT_SYMBOL(__preempt_spin_lock);
4540 void __sched __preempt_write_lock(rwlock_t *lock)
4542 if (preempt_count() > 1) {
4543 _raw_write_lock(lock);
4549 while (rwlock_is_locked(lock))
4552 } while (!_raw_write_trylock(lock));
4555 EXPORT_SYMBOL(__preempt_write_lock);
4556 #endif /* defined(CONFIG_SMP) && defined(CONFIG_PREEMPT) */
4558 #ifdef CONFIG_DELAY_ACCT
4559 int task_running_sys(struct task_struct *p)
4561 return task_running(task_rq(p),p);
4563 EXPORT_SYMBOL(task_running_sys);
4566 #ifdef CONFIG_CKRM_CPU_SCHEDULE
4568 * return the classqueue object of a certain processor
4569 * Note: not supposed to be used in performance sensitive functions
4571 struct classqueue_struct * get_cpu_classqueue(int cpu)
4573 return (& (cpu_rq(cpu)->classqueue) );